Upload 95 files

.gitattributes

@@ -33,3 +33,19 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+OmniParser/imgs/demo_image_som.jpg filter=lfs diff=lfs merge=lfs -text
+OmniParser/imgs/demo_image.jpg filter=lfs diff=lfs merge=lfs -text
+OmniParser/imgs/excel.png filter=lfs diff=lfs merge=lfs -text
+OmniParser/imgs/google_page.png filter=lfs diff=lfs merge=lfs -text
+OmniParser/imgs/header_bar.png filter=lfs diff=lfs merge=lfs -text
+OmniParser/imgs/ios.png filter=lfs diff=lfs merge=lfs -text
+OmniParser/imgs/mobile.png filter=lfs diff=lfs merge=lfs -text
+OmniParser/imgs/onenote.png filter=lfs diff=lfs merge=lfs -text
+OmniParser/imgs/saved_image_demo.png filter=lfs diff=lfs merge=lfs -text
+OmniParser/imgs/som_overlaid_omni.png filter=lfs diff=lfs merge=lfs -text
+OmniParser/imgs/teams.png filter=lfs diff=lfs merge=lfs -text
+OmniParser/imgs/windows_home.png filter=lfs diff=lfs merge=lfs -text
+OmniParser/imgs/windows_multitab.png filter=lfs diff=lfs merge=lfs -text
+OmniParser/imgs/windows_vm.png filter=lfs diff=lfs merge=lfs -text
+OmniParser/imgs/windows.png filter=lfs diff=lfs merge=lfs -text
+OmniParser/imgs/word.png filter=lfs diff=lfs merge=lfs -text

OmniParser/.gitignore

@@ -0,0 +1,13 @@
+weights/icon_caption_blip2
+weights/icon_caption_florence
+weights/icon_detect/
+weights/icon_detect_v1_5/
+weights/icon_detect_v1_5_2/
+.gradio
+__pycache__/
+debug.ipynb
+util/__pycache__/
+index.html?linkid=2289031
+wget-log
+weights/icon_caption_florence_v2/
+omnitool/gradio/uploads/

OmniParser/LICENSE

@@ -0,0 +1,395 @@
+Attribution 4.0 International
+
+=======================================================================
+
+Creative Commons Corporation ("Creative Commons") is not a law firm and
+does not provide legal services or legal advice. Distribution of
+Creative Commons public licenses does not create a lawyer-client or
+other relationship. Creative Commons makes its licenses and related
+information available on an "as-is" basis. Creative Commons gives no
+warranties regarding its licenses, any material licensed under their
+terms and conditions, or any related information. Creative Commons
+disclaims all liability for damages resulting from their use to the
+fullest extent possible.
+
+Using Creative Commons Public Licenses
+
+Creative Commons public licenses provide a standard set of terms and
+conditions that creators and other rights holders may use to share
+original works of authorship and other material subject to copyright
+and certain other rights specified in the public license below. The
+following considerations are for informational purposes only, are not
+exhaustive, and do not form part of our licenses.
+
+     Considerations for licensors: Our public licenses are
+     intended for use by those authorized to give the public
+     permission to use material in ways otherwise restricted by
+     copyright and certain other rights. Our licenses are
+     irrevocable. Licensors should read and understand the terms
+     and conditions of the license they choose before applying it.
+     Licensors should also secure all rights necessary before
+     applying our licenses so that the public can reuse the
+     material as expected. Licensors should clearly mark any
+     material not subject to the license. This includes other CC-
+     licensed material, or material used under an exception or
+     limitation to copyright. More considerations for licensors:
+	wiki.creativecommons.org/Considerations_for_licensors
+
+     Considerations for the public: By using one of our public
+     licenses, a licensor grants the public permission to use the
+     licensed material under specified terms and conditions. If
+     the licensor's permission is not necessary for any reason--for
+     example, because of any applicable exception or limitation to
+     copyright--then that use is not regulated by the license. Our
+     licenses grant only permissions under copyright and certain
+     other rights that a licensor has authority to grant. Use of
+     the licensed material may still be restricted for other
+     reasons, including because others have copyright or other
+     rights in the material. A licensor may make special requests,
+     such as asking that all changes be marked or described.
+     Although not required by our licenses, you are encouraged to
+     respect those requests where reasonable. More_considerations
+     for the public:
+	wiki.creativecommons.org/Considerations_for_licensees
+
+=======================================================================
+
+Creative Commons Attribution 4.0 International Public License
+
+By exercising the Licensed Rights (defined below), You accept and agree
+to be bound by the terms and conditions of this Creative Commons
+Attribution 4.0 International Public License ("Public License"). To the
+extent this Public License may be interpreted as a contract, You are
+granted the Licensed Rights in consideration of Your acceptance of
+these terms and conditions, and the Licensor grants You such rights in
+consideration of benefits the Licensor receives from making the
+Licensed Material available under these terms and conditions.
+
+
+Section 1 -- Definitions.
+
+  a. Adapted Material means material subject to Copyright and Similar
+     Rights that is derived from or based upon the Licensed Material
+     and in which the Licensed Material is translated, altered,
+     arranged, transformed, or otherwise modified in a manner requiring
+     permission under the Copyright and Similar Rights held by the
+     Licensor. For purposes of this Public License, where the Licensed
+     Material is a musical work, performance, or sound recording,
+     Adapted Material is always produced where the Licensed Material is
+     synched in timed relation with a moving image.
+
+  b. Adapter's License means the license You apply to Your Copyright
+     and Similar Rights in Your contributions to Adapted Material in
+     accordance with the terms and conditions of this Public License.
+
+  c. Copyright and Similar Rights means copyright and/or similar rights
+     closely related to copyright including, without limitation,
+     performance, broadcast, sound recording, and Sui Generis Database
+     Rights, without regard to how the rights are labeled or
+     categorized. For purposes of this Public License, the rights
+     specified in Section 2(b)(1)-(2) are not Copyright and Similar
+     Rights.
+
+  d. Effective Technological Measures means those measures that, in the
+     absence of proper authority, may not be circumvented under laws
+     fulfilling obligations under Article 11 of the WIPO Copyright
+     Treaty adopted on December 20, 1996, and/or similar international
+     agreements.
+
+  e. Exceptions and Limitations means fair use, fair dealing, and/or
+     any other exception or limitation to Copyright and Similar Rights
+     that applies to Your use of the Licensed Material.
+
+  f. Licensed Material means the artistic or literary work, database,
+     or other material to which the Licensor applied this Public
+     License.
+
+  g. Licensed Rights means the rights granted to You subject to the
+     terms and conditions of this Public License, which are limited to
+     all Copyright and Similar Rights that apply to Your use of the
+     Licensed Material and that the Licensor has authority to license.
+
+  h. Licensor means the individual(s) or entity(ies) granting rights
+     under this Public License.
+
+  i. Share means to provide material to the public by any means or
+     process that requires permission under the Licensed Rights, such
+     as reproduction, public display, public performance, distribution,
+     dissemination, communication, or importation, and to make material
+     available to the public including in ways that members of the
+     public may access the material from a place and at a time
+     individually chosen by them.
+
+  j. Sui Generis Database Rights means rights other than copyright
+     resulting from Directive 96/9/EC of the European Parliament and of
+     the Council of 11 March 1996 on the legal protection of databases,
+     as amended and/or succeeded, as well as other essentially
+     equivalent rights anywhere in the world.
+
+  k. You means the individual or entity exercising the Licensed Rights
+     under this Public License. Your has a corresponding meaning.
+
+
+Section 2 -- Scope.
+
+  a. License grant.
+
+       1. Subject to the terms and conditions of this Public License,
+          the Licensor hereby grants You a worldwide, royalty-free,
+          non-sublicensable, non-exclusive, irrevocable license to
+          exercise the Licensed Rights in the Licensed Material to:
+
+            a. reproduce and Share the Licensed Material, in whole or
+               in part; and
+
+            b. produce, reproduce, and Share Adapted Material.
+
+       2. Exceptions and Limitations. For the avoidance of doubt, where
+          Exceptions and Limitations apply to Your use, this Public
+          License does not apply, and You do not need to comply with
+          its terms and conditions.
+
+       3. Term. The term of this Public License is specified in Section
+          6(a).
+
+       4. Media and formats; technical modifications allowed. The
+          Licensor authorizes You to exercise the Licensed Rights in
+          all media and formats whether now known or hereafter created,
+          and to make technical modifications necessary to do so. The
+          Licensor waives and/or agrees not to assert any right or
+          authority to forbid You from making technical modifications
+          necessary to exercise the Licensed Rights, including
+          technical modifications necessary to circumvent Effective
+          Technological Measures. For purposes of this Public License,
+          simply making modifications authorized by this Section 2(a)
+          (4) never produces Adapted Material.
+
+       5. Downstream recipients.
+
+            a. Offer from the Licensor -- Licensed Material. Every
+               recipient of the Licensed Material automatically
+               receives an offer from the Licensor to exercise the
+               Licensed Rights under the terms and conditions of this
+               Public License.
+
+            b. No downstream restrictions. You may not offer or impose
+               any additional or different terms or conditions on, or
+               apply any Effective Technological Measures to, the
+               Licensed Material if doing so restricts exercise of the
+               Licensed Rights by any recipient of the Licensed
+               Material.
+
+       6. No endorsement. Nothing in this Public License constitutes or
+          may be construed as permission to assert or imply that You
+          are, or that Your use of the Licensed Material is, connected
+          with, or sponsored, endorsed, or granted official status by,
+          the Licensor or others designated to receive attribution as
+          provided in Section 3(a)(1)(A)(i).
+
+  b. Other rights.
+
+       1. Moral rights, such as the right of integrity, are not
+          licensed under this Public License, nor are publicity,
+          privacy, and/or other similar personality rights; however, to
+          the extent possible, the Licensor waives and/or agrees not to
+          assert any such rights held by the Licensor to the limited
+          extent necessary to allow You to exercise the Licensed
+          Rights, but not otherwise.
+
+       2. Patent and trademark rights are not licensed under this
+          Public License.
+
+       3. To the extent possible, the Licensor waives any right to
+          collect royalties from You for the exercise of the Licensed
+          Rights, whether directly or through a collecting society
+          under any voluntary or waivable statutory or compulsory
+          licensing scheme. In all other cases the Licensor expressly
+          reserves any right to collect such royalties.
+
+
+Section 3 -- License Conditions.
+
+Your exercise of the Licensed Rights is expressly made subject to the
+following conditions.
+
+  a. Attribution.
+
+       1. If You Share the Licensed Material (including in modified
+          form), You must:
+
+            a. retain the following if it is supplied by the Licensor
+               with the Licensed Material:
+
+                 i. identification of the creator(s) of the Licensed
+                    Material and any others designated to receive
+                    attribution, in any reasonable manner requested by
+                    the Licensor (including by pseudonym if
+                    designated);
+
+                ii. a copyright notice;
+
+               iii. a notice that refers to this Public License;
+
+                iv. a notice that refers to the disclaimer of
+                    warranties;
+
+                 v. a URI or hyperlink to the Licensed Material to the
+                    extent reasonably practicable;
+
+            b. indicate if You modified the Licensed Material and
+               retain an indication of any previous modifications; and
+
+            c. indicate the Licensed Material is licensed under this
+               Public License, and include the text of, or the URI or
+               hyperlink to, this Public License.
+
+       2. You may satisfy the conditions in Section 3(a)(1) in any
+          reasonable manner based on the medium, means, and context in
+          which You Share the Licensed Material. For example, it may be
+          reasonable to satisfy the conditions by providing a URI or
+          hyperlink to a resource that includes the required
+          information.
+
+       3. If requested by the Licensor, You must remove any of the
+          information required by Section 3(a)(1)(A) to the extent
+          reasonably practicable.
+
+       4. If You Share Adapted Material You produce, the Adapter's
+          License You apply must not prevent recipients of the Adapted
+          Material from complying with this Public License.
+
+
+Section 4 -- Sui Generis Database Rights.
+
+Where the Licensed Rights include Sui Generis Database Rights that
+apply to Your use of the Licensed Material:
+
+  a. for the avoidance of doubt, Section 2(a)(1) grants You the right
+     to extract, reuse, reproduce, and Share all or a substantial
+     portion of the contents of the database;
+
+  b. if You include all or a substantial portion of the database
+     contents in a database in which You have Sui Generis Database
+     Rights, then the database in which You have Sui Generis Database
+     Rights (but not its individual contents) is Adapted Material; and
+
+  c. You must comply with the conditions in Section 3(a) if You Share
+     all or a substantial portion of the contents of the database.
+
+For the avoidance of doubt, this Section 4 supplements and does not
+replace Your obligations under this Public License where the Licensed
+Rights include other Copyright and Similar Rights.
+
+
+Section 5 -- Disclaimer of Warranties and Limitation of Liability.
+
+  a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
+     EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
+     AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
+     ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
+     IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
+     WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
+     PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
+     ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
+     KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
+     ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.
+
+  b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
+     TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
+     NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
+     INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
+     COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
+     USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
+     ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
+     DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
+     IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.
+
+  c. The disclaimer of warranties and limitation of liability provided
+     above shall be interpreted in a manner that, to the extent
+     possible, most closely approximates an absolute disclaimer and
+     waiver of all liability.
+
+
+Section 6 -- Term and Termination.
+
+  a. This Public License applies for the term of the Copyright and
+     Similar Rights licensed here. However, if You fail to comply with
+     this Public License, then Your rights under this Public License
+     terminate automatically.
+
+  b. Where Your right to use the Licensed Material has terminated under
+     Section 6(a), it reinstates:
+
+       1. automatically as of the date the violation is cured, provided
+          it is cured within 30 days of Your discovery of the
+          violation; or
+
+       2. upon express reinstatement by the Licensor.
+
+     For the avoidance of doubt, this Section 6(b) does not affect any
+     right the Licensor may have to seek remedies for Your violations
+     of this Public License.
+
+  c. For the avoidance of doubt, the Licensor may also offer the
+     Licensed Material under separate terms or conditions or stop
+     distributing the Licensed Material at any time; however, doing so
+     will not terminate this Public License.
+
+  d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
+     License.
+
+
+Section 7 -- Other Terms and Conditions.
+
+  a. The Licensor shall not be bound by any additional or different
+     terms or conditions communicated by You unless expressly agreed.
+
+  b. Any arrangements, understandings, or agreements regarding the
+     Licensed Material not stated herein are separate from and
+     independent of the terms and conditions of this Public License.
+
+
+Section 8 -- Interpretation.
+
+  a. For the avoidance of doubt, this Public License does not, and
+     shall not be interpreted to, reduce, limit, restrict, or impose
+     conditions on any use of the Licensed Material that could lawfully
+     be made without permission under this Public License.
+
+  b. To the extent possible, if any provision of this Public License is
+     deemed unenforceable, it shall be automatically reformed to the
+     minimum extent necessary to make it enforceable. If the provision
+     cannot be reformed, it shall be severed from this Public License
+     without affecting the enforceability of the remaining terms and
+     conditions.
+
+  c. No term or condition of this Public License will be waived and no
+     failure to comply consented to unless expressly agreed to by the
+     Licensor.
+
+  d. Nothing in this Public License constitutes or may be interpreted
+     as a limitation upon, or waiver of, any privileges and immunities
+     that apply to the Licensor or You, including from the legal
+     processes of any jurisdiction or authority.
+
+
+=======================================================================
+
+Creative Commons is not a party to its public
+licenses. Notwithstanding, Creative Commons may elect to apply one of
+its public licenses to material it publishes and in those instances
+will be considered the “Licensor.” The text of the Creative Commons
+public licenses is dedicated to the public domain under the CC0 Public
+Domain Dedication. Except for the limited purpose of indicating that
+material is shared under a Creative Commons public license or as
+otherwise permitted by the Creative Commons policies published at
+creativecommons.org/policies, Creative Commons does not authorize the
+use of the trademark "Creative Commons" or any other trademark or logo
+of Creative Commons without its prior written consent including,
+without limitation, in connection with any unauthorized modifications
+to any of its public licenses or any other arrangements,
+understandings, or agreements concerning use of licensed material. For
+the avoidance of doubt, this paragraph does not form part of the
+public licenses.
+
+Creative Commons may be contacted at creativecommons.org.

OmniParser/README.md

@@ -0,0 +1,80 @@
+# OmniParser: Screen Parsing tool for Pure Vision Based GUI Agent
+
+<p align="center">
+  <img src="imgs/logo.png" alt="Logo">
+</p>
+<!-- <a href="https://trendshift.io/repositories/12975" target="_blank"><img src="https://trendshift.io/api/badge/repositories/12975" alt="microsoft%2FOmniParser | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a> -->
+
+[![arXiv](https://img.shields.io/badge/Paper-green)](https://arxiv.org/abs/2408.00203)
+[![License](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
+
+📢 [[Project Page](https://microsoft.github.io/OmniParser/)] [[V2 Blog Post](https://www.microsoft.com/en-us/research/articles/omniparser-v2-turning-any-llm-into-a-computer-use-agent/)] [[Models V2](https://huggingface.co/microsoft/OmniParser-v2.0)] [[Models V1.5](https://huggingface.co/microsoft/OmniParser)] [[HuggingFace Space Demo](https://huggingface.co/spaces/microsoft/OmniParser-v2)]
+
+**OmniParser** is a comprehensive method for parsing user interface screenshots into structured and easy-to-understand elements, which significantly enhances the ability of GPT-4V to generate actions that can be accurately grounded in the corresponding regions of the interface. 
+
+## News
+- [2025/3] We support local logging of trajecotry so that you can use OmniParser+OmniTool to build training data pipeline for your favorate agent in your domain. [Documentation WIP]
+- [2025/3] We are gradually adding multi agents orchstration and improving user interface in OmniTool for better experience.
+- [2025/2] We release OmniParser V2 [checkpoints](https://huggingface.co/microsoft/OmniParser-v2.0). [Watch Video](https://1drv.ms/v/c/650b027c18d5a573/EWXbVESKWo9Buu6OYCwg06wBeoM97C6EOTG6RjvWLEN1Qg?e=alnHGC)
+- [2025/2] We introduce OmniTool: Control a Windows 11 VM with OmniParser + your vision model of choice. OmniTool supports out of the box the following large language models - OpenAI (4o/o1/o3-mini), DeepSeek (R1), Qwen (2.5VL) or Anthropic Computer Use. [Watch Video](https://1drv.ms/v/c/650b027c18d5a573/EehZ7RzY69ZHn-MeQHrnnR4BCj3by-cLLpUVlxMjF4O65Q?e=8LxMgX)
+- [2025/1] V2 is coming. We achieve new state of the art results 39.5% on the new grounding benchmark [Screen Spot Pro](https://github.com/likaixin2000/ScreenSpot-Pro-GUI-Grounding/tree/main) with OmniParser v2 (will be released soon)! Read more details [here](https://github.com/microsoft/OmniParser/tree/master/docs/Evaluation.md).
+- [2024/11] We release an updated version, OmniParser V1.5 which features 1) more fine grained/small icon detection, 2) prediction of whether each screen element is interactable or not. Examples in the demo.ipynb. 
+- [2024/10] OmniParser was the #1 trending model on huggingface model hub (starting 10/29/2024). 
+- [2024/10] Feel free to checkout our demo on [huggingface space](https://huggingface.co/spaces/microsoft/OmniParser)! (stay tuned for OmniParser + Claude Computer Use)
+- [2024/10] Both Interactive Region Detection Model and Icon functional description model are released! [Hugginface models](https://huggingface.co/microsoft/OmniParser)
+- [2024/09] OmniParser achieves the best performance on [Windows Agent Arena](https://microsoft.github.io/WindowsAgentArena/)! 
+
+## Install 
+First clone the repo, and then install environment:
+```python
+cd OmniParser
+conda create -n "omni" python==3.12
+conda activate omni
+pip install -r requirements.txt
+```
+
+Ensure you have the V2 weights downloaded in weights folder (ensure caption weights folder is called icon_caption_florence). If not download them with:
+```
+   # download the model checkpoints to local directory OmniParser/weights/
+   for f in icon_detect/{train_args.yaml,model.pt,model.yaml} icon_caption/{config.json,generation_config.json,model.safetensors}; do huggingface-cli download microsoft/OmniParser-v2.0 "$f" --local-dir weights; done
+   mv weights/icon_caption weights/icon_caption_florence
+```
+
+<!-- ## [deprecated]
+Then download the model ckpts files in: https://huggingface.co/microsoft/OmniParser, and put them under weights/, default folder structure is: weights/icon_detect, weights/icon_caption_florence, weights/icon_caption_blip2. 
+
+For v1: 
+convert the safetensor to .pt file. 
+```python
+python weights/convert_safetensor_to_pt.py
+
+For v1.5: 
+download 'model_v1_5.pt' from https://huggingface.co/microsoft/OmniParser/tree/main/icon_detect_v1_5, make a new dir: weights/icon_detect_v1_5, and put it inside the folder. No weight conversion is needed. 
+``` -->
+
+## Examples:
+We put together a few simple examples in the demo.ipynb. 
+
+## Gradio Demo
+To run gradio demo, simply run:
+```python
+python gradio_demo.py
+```
+
+## Model Weights License
+For the model checkpoints on huggingface model hub, please note that icon_detect model is under AGPL license since it is a license inherited from the original yolo model. And icon_caption_blip2 & icon_caption_florence is under MIT license. Please refer to the LICENSE file in the folder of each model: https://huggingface.co/microsoft/OmniParser.
+
+## 📚 Citation
+Our technical report can be found [here](https://arxiv.org/abs/2408.00203).
+If you find our work useful, please consider citing our work:
+```
+@misc{lu2024omniparserpurevisionbased,
+      title={OmniParser for Pure Vision Based GUI Agent}, 
+      author={Yadong Lu and Jianwei Yang and Yelong Shen and Ahmed Awadallah},
+      year={2024},
+      eprint={2408.00203},
+      archivePrefix={arXiv},
+      primaryClass={cs.CV},
+      url={https://arxiv.org/abs/2408.00203}, 
+}
+```

OmniParser/SECURITY.md

@@ -0,0 +1,41 @@
+<!-- BEGIN MICROSOFT SECURITY.MD V0.0.9 BLOCK -->
+
+## Security
+
+Microsoft takes the security of our software products and services seriously, which includes all source code repositories managed through our GitHub organizations, which include [Microsoft](https://github.com/Microsoft), [Azure](https://github.com/Azure), [DotNet](https://github.com/dotnet), [AspNet](https://github.com/aspnet) and [Xamarin](https://github.com/xamarin).
+
+If you believe you have found a security vulnerability in any Microsoft-owned repository that meets [Microsoft's definition of a security vulnerability](https://aka.ms/security.md/definition), please report it to us as described below.
+
+## Reporting Security Issues
+
+**Please do not report security vulnerabilities through public GitHub issues.**
+
+Instead, please report them to the Microsoft Security Response Center (MSRC) at [https://msrc.microsoft.com/create-report](https://aka.ms/security.md/msrc/create-report).
+
+If you prefer to submit without logging in, send email to [secure@microsoft.com](mailto:secure@microsoft.com).  If possible, encrypt your message with our PGP key; please download it from the [Microsoft Security Response Center PGP Key page](https://aka.ms/security.md/msrc/pgp).
+
+You should receive a response within 24 hours. If for some reason you do not, please follow up via email to ensure we received your original message. Additional information can be found at [microsoft.com/msrc](https://www.microsoft.com/msrc). 
+
+Please include the requested information listed below (as much as you can provide) to help us better understand the nature and scope of the possible issue:
+
+  * Type of issue (e.g. buffer overflow, SQL injection, cross-site scripting, etc.)
+  * Full paths of source file(s) related to the manifestation of the issue
+  * The location of the affected source code (tag/branch/commit or direct URL)
+  * Any special configuration required to reproduce the issue
+  * Step-by-step instructions to reproduce the issue
+  * Proof-of-concept or exploit code (if possible)
+  * Impact of the issue, including how an attacker might exploit the issue
+
+This information will help us triage your report more quickly.
+
+If you are reporting for a bug bounty, more complete reports can contribute to a higher bounty award. Please visit our [Microsoft Bug Bounty Program](https://aka.ms/security.md/msrc/bounty) page for more details about our active programs.
+
+## Preferred Languages
+
+We prefer all communications to be in English.
+
+## Policy
+
+Microsoft follows the principle of [Coordinated Vulnerability Disclosure](https://aka.ms/security.md/cvd).
+
+<!-- END MICROSOFT SECURITY.MD BLOCK -->

OmniParser/docs/Evaluation.md

@@ -0,0 +1,4 @@
+# Eval setup for ScreenSpot Pro
+We adapt the eval code from ScreenSpot Pro (ss pro) official [repo](https://github.com/likaixin2000/ScreenSpot-Pro-GUI-Grounding/tree/main). This folder contains the inference script/results on this benchmark. We going through legal review proces to release omniparser v2. Once it is done, we will update the file so that it can load the v2 model. 
+1. eval/ss_pro_gpt4o_omniv2.py: contains the prompt we use, it can be dropped in replacement for this [file](https://github.com/likaixin2000/ScreenSpot-Pro-GUI-Grounding/blob/main/models/gpt4x.py) in the original ss pro repo.
+2. eval/logs_sspro_omniv2.json: contains the inferenced results for ss pro using GPT4o+OmniParserv2. 

OmniParser/eval/ss_pro_gpt4o_omniv2.py

@@ -0,0 +1,411 @@
+import os
+import re
+import ast
+import base64
+from io import BytesIO
+from PIL import Image
+from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
+
+import openai
+from openai import BadRequestError
+
+model_name = "gpt-4o-2024-05-13"
+OPENAI_KEY = os.environ.get("OPENAI_API_KEY")
+
+def convert_pil_image_to_base64(image):
+    buffered = BytesIO()
+    image.save(buffered, format="PNG")
+    return base64.b64encode(buffered.getvalue()).decode()
+
+
+from models.utils import get_som_labeled_img, check_ocr_box, get_caption_model_processor, get_yolo_model
+import torch
+from ultralytics import YOLO
+from PIL import Image
+device = 'cuda' if torch.cuda.is_available() else 'cpu'
+SOM_MODEL_PATH='...'
+CAPTION_MODEL_PATH='...'
+som_model = get_yolo_model(SOM_MODEL_PATH)
+
+som_model.to(device)
+print('model to {}'.format(device))
+
+# two choices for caption model: fine-tuned blip2 or florence2
+import importlib
+caption_model_processor = get_caption_model_processor(model_name="florence2", model_name_or_path="CAPTION_MODEL_PATH", device=device)
+
+def omniparser_parse(image, image_path):
+    box_overlay_ratio = max(image.size) / 3200
+    draw_bbox_config = {
+        'text_scale': 0.8 * box_overlay_ratio,
+        'text_thickness': max(int(2 * box_overlay_ratio), 1),
+        'text_padding': max(int(3 * box_overlay_ratio), 1),
+        'thickness': max(int(3 * box_overlay_ratio), 1),
+    }
+    BOX_TRESHOLD = 0.05
+
+    ocr_bbox_rslt, is_goal_filtered = check_ocr_box(image_path, display_img = False, output_bb_format='xyxy', goal_filtering=None, easyocr_args={'paragraph': False, 'text_threshold':0.5, 'canvas_size':max(image.size), 'decoder':'beamsearch', 'beamWidth':10, 'batch_size':256}, use_paddleocr=False)
+    text, ocr_bbox = ocr_bbox_rslt
+
+    dino_labled_img, label_coordinates, parsed_content_list = get_som_labeled_img(image_path, som_model, BOX_TRESHOLD = BOX_TRESHOLD, output_coord_in_ratio=True, ocr_bbox=ocr_bbox,draw_bbox_config=draw_bbox_config, caption_model_processor=caption_model_processor, ocr_text=text,use_local_semantics=True, iou_threshold=0.7, scale_img=False, batch_size=128)
+    return dino_labled_img, label_coordinates, parsed_content_list
+    
+def reformat_messages(parsed_content_list):
+    screen_info = ""
+    for idx, element in enumerate(parsed_content_list):
+        element['idx'] = idx
+        if element['type'] == 'text':
+            screen_info += f'''<p id={idx} class="text" alt="{element['content']}"> </p>\n'''
+            # screen_info += f'ID: {idx}, Text: {element["content"]}\n'
+        elif element['type'] == 'icon':
+            screen_info += f'''<img id={idx} class="icon" alt="{element['content']}"> </img>\n'''
+            # screen_info += f'ID: {idx}, Icon: {element["content"]}\n'
+    return screen_info
+
+PROMPT_TEMPLATE_SEECLICK_PARSED_CONTENT = '''Please generate the next move according to the UI screenshot and task instruction. You will be presented with a screenshot image. Also you will be given each bounding box's description in a list. To complete the task, You should choose a related bbox to click based on the bbox descriptions. 
+Task instruction: {}. 
+Here is the list of all detected bounding boxes by IDs and their descriptions: {}. Keep in mind the description for Text Boxes are likely more accurate than the description for Icon Boxes. 
+Requirement: 1. You should first give a reasonable description of the current screenshot, and give a short analysis of how can the user task be achieved. 2. Then make an educated guess of bbox id to click in order to complete the task based on the bounding boxes descriptions. 3. Your answer should follow the following format: {{"Analysis": xxx, "Click BBox ID": "y"}}. Do not include any other info. Some examples: {}. The task is to {}. Retrieve the bbox id where its description matches the task instruction. Now start your answer:'''
+
+# PROMPT_TEMPLATE_SEECLICK_PARSED_CONTENT_v1 = "The instruction is to {}. \nHere is the list of all detected bounding boxes by IDs and their descriptions: {}. \nKeep in mind the description for Text Boxes are likely more accurate than the description for Icon Boxes. \n Requirement: 1. You should first give a reasonable description of the current screenshot, and give a step by step analysis of how can the user task be achieved. 2. Then make an educated guess of bbox id to click in order to complete the task using both the visual information from the screenshot image and the bounding boxes descriptions. 3. Your answer should follow the following format: {{'Analysis': 'xxx', 'Click BBox ID': 'y'}}. Please do not include any other info."
+PROMPT_TEMPLATE_SEECLICK_PARSED_CONTENT_v1 = "The instruction is to {}. \nHere is the list of all detected bounding boxes by IDs and their descriptions: {}. \nKeep in mind the description for Text Boxes are likely more accurate than the description for Icon Boxes. \n Requirement: 1. You should first give a reasonable description of the current screenshot, and give a some analysis of how can the user instruction be achieved by a single click. 2. Then make an educated guess of bbox id to click in order to complete the task using both the visual information from the screenshot image and the bounding boxes descriptions. REMEMBER: the task instruction must be achieved by one single click. 3. Your answer should follow the following format: {{'Analysis': 'xxx', 'Click BBox ID': 'y'}}. Please do not include any other info."
+
+
+FEWSHOT_EXAMPLE = '''Example 1: Task instruction: Next page. \n{"Analysis": "Based on the screenshot and icon descriptions, I should click on the next page icon, which is labeled with box ID x in the bounding box list", "Click BBox ID": "x"}\n\n
+Example 2: Task instruction: Search on google. \n{"Analysis": "Based on the screenshot and icon descriptions, I should click on the 'Search' box, which is labeled with box ID y in the bounding box list", "Click BBox ID": "y"}'''
+
+
+
+
+from azure.identity import AzureCliCredential, DefaultAzureCredential, get_bearer_token_provider
+from openai import AzureOpenAI
+from models.utils import get_pred_phi3v, extract_dict_from_text, get_phi3v_model_dict
+
+class GPT4XModel():
+    def __init__(self, model_name="gpt-4o-2024-05-13", use_managed_identity=False):
+        self.client = openai.OpenAI(
+            api_key=OPENAI_KEY,
+        )
+        self.model_name = model_name
+        if model_name == 'phi35v':
+            self.model_dict = get_phi3v_model_dict()
+
+    def load_model(self):
+        pass
+    
+    def set_generation_config(self, **kwargs):
+        self.override_generation_config.update(kwargs)
+
+    def ground_only_positive_phi35v(self, instruction, image):
+        if isinstance(image, str):
+            image_path = image
+            assert os.path.exists(image_path) and os.path.isfile(image_path), "Invalid input image path."
+            image = Image.open(image_path).convert('RGB')
+        assert isinstance(image, Image.Image), "Invalid input image."
+        
+        base64_image = convert_pil_image_to_base64(image)
+        dino_labled_img, label_coordinates, parsed_content_list = omniparser_parse(image, image_path)
+        screen_info = reformat_messages(parsed_content_list)
+        prompt_origin = PROMPT_TEMPLATE_SEECLICK_PARSED_CONTENT.format(instruction, screen_info, FEWSHOT_EXAMPLE, instruction)
+        # prompt_origin = PROMPT_TEMPLATE_SEECLICK_PARSED_CONTENT_v1.format(instruction, screen_info)
+
+        # Use the get_pred_phi3v function to get predictions
+        icon_id, bbox, click_point, response_text = get_pred_phi3v(prompt_origin, (base64_image, dino_labled_img), label_coordinates, id_key='Click ID', model_dict=self.model_dict)
+
+        result_dict = {
+            "result": "positive",
+            "bbox": bbox,
+            "point": click_point,
+            "raw_response": response_text,
+            'dino_labled_img': dino_labled_img,
+            'screen_info': screen_info,
+        }
+        
+        return result_dict
+
+    def ground_only_positive(self, instruction, image):
+        if isinstance(image, str):
+            image_path = image
+            assert os.path.exists(image_path) and os.path.isfile(image_path), "Invalid input image path."
+            image = Image.open(image_path).convert('RGB')
+        assert isinstance(image, Image.Image), "Invalid input image."
+        
+        base64_image = convert_pil_image_to_base64(image)
+        dino_labled_img, label_coordinates, parsed_content_list = omniparser_parse(image, image_path)
+        screen_info = reformat_messages(parsed_content_list)
+        # prompt_origin = PROMPT_TEMPLATE_SEECLICK_PARSED_CONTENT.format(screen_info, FEWSHOT_EXAMPLE, instruction)
+        prompt_origin = PROMPT_TEMPLATE_SEECLICK_PARSED_CONTENT_v1.format(instruction, screen_info)
+
+        try:
+            response = self.client.chat.completions.create(
+                model=self.model_name,
+                messages=[
+                    {
+                        "role": "system",
+                        "content": [
+                            # {"type": "text", "text": "You are an expert in using electronic devices and interacting with graphic interfaces. You should not call any external tools."}
+                            {"type": "text", "text": '''You are an expert at completing instructions on GUI screens. 
+               You will be presented with two images. The first is the original screenshot. The second is the same screenshot with some numeric tags. You will also be provided with some descriptions of the bbox, and your task is to choose the numeric bbox idx you want to click in order to complete the user instruction.'''}
+                        ],
+                    },
+                    {
+                        "role": "user",
+                        "content": [
+                            {
+                                "type": "text", 
+                                "text": prompt_origin
+
+                            },
+                            {
+                                "type": "image_url",
+                                "image_url": {
+                                    "url": f"data:image/png;base64,{base64_image}",
+                                }
+                            },
+                            {
+                                "type": "image_url",
+                                "image_url": {
+                                    "url": f"data:image/png;base64,{dino_labled_img}",
+                                }
+                            },
+                        ],
+                    }
+                ],
+                temperature=self.override_generation_config['temperature'],
+                max_tokens=2048,
+            )
+            response_text = response.choices[0].message.content
+        except BadRequestError as e:
+            print("OpenAI BadRequestError:", e)
+            return None
+
+        # Extract bounding box
+        # print("------")
+        # print(grounding_prompt)
+        print("------")
+        print(response_text)
+        # print("------")
+        # Try getting groundings
+        # bbox = extract_first_bounding_box(response_text)
+        # click_point = extract_first_point(response_text)
+        
+        # if not click_point and bbox:
+        #     click_point = [(bbox[0] + bbox[2]) / 2, (bbox[1] + bbox[3]) / 2]
+        response_text = response_text.replace('```json', '').replace('```', '') #TODO: fix this
+
+        try:
+            response_text = ast.literal_eval(response_text)
+            
+            icon_id = response_text['Click BBox ID']
+            bbox = label_coordinates[str(icon_id)]
+            click_point = [bbox[0] + bbox[2]/2, bbox[1] + bbox[3]/2]
+        except:
+            print('error parsing, use regex to parse!!!')
+            response_text = extract_dict_from_text(response_text)
+            icon_id = response_text['Click BBox ID']
+            bbox = label_coordinates[str(icon_id)]
+            click_point = [bbox[0] + bbox[2]/2, bbox[1] + bbox[3]/2]
+
+        result_dict = {
+            "result": "positive",
+            "bbox": bbox,
+            "point": click_point,
+            "raw_response": response_text,
+            'dino_labled_img': dino_labled_img,
+            'screen_info': screen_info,
+        }
+        
+        return result_dict
+
+    def ground_allow_negative(self, instruction, image=None):
+        if isinstance(image, str):
+            image_path = image
+            assert os.path.exists(image_path) and os.path.isfile(image_path), "Invalid input image path."
+            image = Image.open(image_path).convert('RGB')
+        assert isinstance(image, Image.Image), "Invalid input image."
+        
+        base64_image = convert_pil_image_to_base64(image)
+
+        try:
+            response = self.client.chat.completions.create(
+                model=self.model_name,
+                messages=[
+                    {
+                        "role": "system",
+                        "content": [
+                            {"type": "text", "text": "You are an expert in using electronic devices and interacting with graphic interfaces. You should not call any external tools."}
+                        ],
+                    },
+                    {
+                        "role": "user",
+                        "content": [
+                            {
+                                "type": "image_url",
+                                "image_url": {
+                                    "url": f"data:image/png;base64,{base64_image}",
+                                }
+                            },
+                            {
+                                "type": "text", 
+                                "text": "You are asked to find the bounding box of an UI element in the given screenshot corresponding to a given instruction.\n"
+                                        "Don't output any analysis. Output your result in the format of [[x0,y0,x1,y1]], with x and y ranging from 0 to 1. \n"
+                                        "If such element does not exist, output only the text 'Target not existent'.\n"
+                                        "The instruction is:\n"
+                                        f"{instruction}\n"
+                            }
+                        ],
+                    }
+                ],
+                temperature=self.override_generation_config['temperature'],
+                max_tokens=2048,
+            )
+            response_text = response.choices[0].message.content
+        except BadRequestError as e:
+            print("OpenAI BadRequestError:", e)
+            return {
+                "result": "failed"
+            }
+
+        # Extract bounding box
+        # print("------")
+        # print(grounding_prompt)
+        print("------")
+        print(response_text)
+        # print("------")
+
+        if "not existent" in response_text.lower():
+            return {
+                "result": "negative",
+                "bbox": None,
+                "point": None,
+                "raw_response": response_text
+            }
+        
+        # Try getting groundings
+        bbox = extract_first_bounding_box(response_text)
+        click_point = extract_first_point(response_text)
+        
+        if not click_point and bbox:
+            click_point = [(bbox[0] + bbox[2]) / 2, (bbox[1] + bbox[3]) / 2]
+
+        result_dict = {
+            "result": "positive" if bbox or click_point else "negative",
+            "bbox": bbox,
+            "point": click_point,
+            "raw_response": response_text
+        }
+        
+        return result_dict
+
+    
+    def ground_with_uncertainty(self, instruction, image=None):
+        if isinstance(image, str):
+            image_path = image
+            assert os.path.exists(image_path) and os.path.isfile(image_path), "Invalid input image path."
+            image = Image.open(image_path).convert('RGB')
+        assert isinstance(image, Image.Image), "Invalid input image."
+        
+        base64_image = convert_pil_image_to_base64(image)
+
+        try:
+            response = self.client.chat.completions.create(
+                model=self.model_name,
+                messages=[
+                    {
+                        "role": "system",
+                        "content": [
+                            {"type": "text", "text": "You are an expert in using electronic devices and interacting with graphic interfaces. You should not call any external tools."}
+                        ],
+                    },
+                    {
+                        "role": "user",
+                        "content": [
+                            {
+                                "type": "image_url",
+                                "image_url": {
+                                    "url": f"data:image/png;base64,{base64_image}",
+                                }
+                            },
+                            {
+                                "type": "text", 
+                                "text": "You are asked to find the bounding box of an UI element in the given screenshot corresponding to a given instruction.\n"
+                                        "- If such element does not exist in the screenshot, output only the text 'Target not existent'."
+
+                                        "- If you are sure such element exists and you are confident in finding it, output your result in the format of [[x0,y0,x1,y1]], with x and y ranging from 0 to 1. \n"
+                                        "Please find out the bounding box of the UI element corresponding to the following instruction: \n"
+                                        "The instruction is:\n"
+                                        f"{instruction}\n"
+                                        
+                            }
+                        ],
+                    }
+                ],
+                temperature=self.override_generation_config['temperature'],
+                max_tokens=2048,
+            )
+            response_text = response.choices[0].message.content
+        except BadRequestError as e:
+            print("OpenAI BadRequestError:", e)
+            return {
+                "result": "failed"
+            }
+
+        # Extract bounding box
+        # print("------")
+        # print(grounding_prompt)
+        print("------")
+        print(response_text)
+        # print("------")
+
+        if "not found" in response_text.lower():
+            return {
+                "result": "negative",
+                "bbox": None,
+                "point": None,
+                "raw_response": response_text
+            }
+        
+        # Try getting groundings
+        bbox = extract_first_bounding_box(response_text)
+        click_point = extract_first_point(response_text)
+        
+        if not click_point and bbox:
+            click_point = [(bbox[0] + bbox[2]) / 2, (bbox[1] + bbox[3]) / 2]
+
+        result_dict = {
+            "result": "positive",
+            "bbox": bbox,
+            "point": click_point,
+            "raw_response": response_text
+        }
+        
+        return result_dict
+
+def extract_first_bounding_box(text):
+    # Regular expression pattern to match the first bounding box in the format [[x0,y0,x1,y1]]
+    # This captures the entire float value using \d for digits and optional decimal points
+    pattern = r"\[\[(\d+\.\d+|\d+),(\d+\.\d+|\d+),(\d+\.\d+|\d+),(\d+\.\d+|\d+)\]\]"
+    
+    # Search for the first match in the text
+    match = re.search(pattern, text, re.DOTALL)
+    
+    if match:
+        # Capture the bounding box coordinates as floats
+        bbox = [float(match.group(1)), float(match.group(2)), float(match.group(3)), float(match.group(4))]
+        return bbox
+    return None
+
+
+def extract_first_point(text):
+    # Regular expression pattern to match the first point in the format [[x0,y0]]
+    # This captures the entire float value using \d for digits and optional decimal points
+    pattern = r"\[\[(\d+\.\d+|\d+),(\d+\.\d+|\d+)\]\]"
+    
+    # Search for the first match in the text
+    match = re.search(pattern, text, re.DOTALL)
+    
+    if match:
+        point = [float(match.group(1)), float(match.group(2))]
+        return point
+    
+    return None

OmniParser/gradio_demo.py

@@ -0,0 +1,98 @@
+from typing import Optional
+
+import gradio as gr
+import numpy as np
+import torch
+from PIL import Image
+import io
+
+
+import base64, os
+from util.utils import check_ocr_box, get_yolo_model, get_caption_model_processor, get_som_labeled_img
+import torch
+from PIL import Image
+
+yolo_model = get_yolo_model(model_path='weights/icon_detect/model.pt')
+caption_model_processor = get_caption_model_processor(model_name="florence2", model_name_or_path="weights/icon_caption_florence")
+# caption_model_processor = get_caption_model_processor(model_name="blip2", model_name_or_path="weights/icon_caption_blip2")
+
+MARKDOWN = """
+# OmniParser for Pure Vision Based General GUI Agent 🔥
+<div>
+    <a href="https://arxiv.org/pdf/2408.00203">
+        <img src="https://img.shields.io/badge/arXiv-2408.00203-b31b1b.svg" alt="Arxiv" style="display:inline-block;">
+    </a>
+</div>
+
+OmniParser is a screen parsing tool to convert general GUI screen to structured elements. 
+"""
+
+DEVICE = torch.device('cuda')
+
+# @spaces.GPU
+# @torch.inference_mode()
+# @torch.autocast(device_type="cuda", dtype=torch.bfloat16)
+def process(
+    image_input,
+    box_threshold,
+    iou_threshold,
+    use_paddleocr,
+    imgsz
+) -> Optional[Image.Image]:
+
+    box_overlay_ratio = image_input.size[0] / 3200
+    draw_bbox_config = {
+        'text_scale': 0.8 * box_overlay_ratio,
+        'text_thickness': max(int(2 * box_overlay_ratio), 1),
+        'text_padding': max(int(3 * box_overlay_ratio), 1),
+        'thickness': max(int(3 * box_overlay_ratio), 1),
+    }
+    # import pdb; pdb.set_trace()
+
+    ocr_bbox_rslt, is_goal_filtered = check_ocr_box(image_input, display_img = False, output_bb_format='xyxy', goal_filtering=None, easyocr_args={'paragraph': False, 'text_threshold':0.9}, use_paddleocr=use_paddleocr)
+    text, ocr_bbox = ocr_bbox_rslt
+    dino_labled_img, label_coordinates, parsed_content_list = get_som_labeled_img(image_input, yolo_model, BOX_TRESHOLD = box_threshold, output_coord_in_ratio=True, ocr_bbox=ocr_bbox,draw_bbox_config=draw_bbox_config, caption_model_processor=caption_model_processor, ocr_text=text,iou_threshold=iou_threshold, imgsz=imgsz,)  
+    image = Image.open(io.BytesIO(base64.b64decode(dino_labled_img)))
+    print('finish processing')
+    parsed_content_list = '\n'.join([f'icon {i}: ' + str(v) for i,v in enumerate(parsed_content_list)])
+    # parsed_content_list = str(parsed_content_list)
+    return image, str(parsed_content_list)
+
+with gr.Blocks() as demo:
+    gr.Markdown(MARKDOWN)
+    with gr.Row():
+        with gr.Column():
+            image_input_component = gr.Image(
+                type='pil', label='Upload image')
+            # set the threshold for removing the bounding boxes with low confidence, default is 0.05
+            box_threshold_component = gr.Slider(
+                label='Box Threshold', minimum=0.01, maximum=1.0, step=0.01, value=0.05)
+            # set the threshold for removing the bounding boxes with large overlap, default is 0.1
+            iou_threshold_component = gr.Slider(
+                label='IOU Threshold', minimum=0.01, maximum=1.0, step=0.01, value=0.1)
+            use_paddleocr_component = gr.Checkbox(
+                label='Use PaddleOCR', value=True)
+            imgsz_component = gr.Slider(
+                label='Icon Detect Image Size', minimum=640, maximum=1920, step=32, value=640)
+            submit_button_component = gr.Button(
+                value='Submit', variant='primary')
+        with gr.Column():
+            image_output_component = gr.Image(type='pil', label='Image Output')
+            text_output_component = gr.Textbox(label='Parsed screen elements', placeholder='Text Output')
+
+    submit_button_component.click(
+        fn=process,
+        inputs=[
+            image_input_component,
+            box_threshold_component,
+            iou_threshold_component,
+            use_paddleocr_component,
+            imgsz_component
+        ],
+        outputs=[image_output_component, text_output_component]
+    )
+
+#demo.launch(debug=True, show_error=True, share=True)
+#demo.launch(share=True, server_port=7861, server_name='0.0.0.0')
+#demo.launch(share=True, server_port=7862, server_name='127.0.0.1')
+demo.launch(server_name="0.0.0.0", server_port=7860, debug=True, show_error=True, share=True)

OmniParser/omnitool/gradio/.gitignore

@@ -0,0 +1 @@
+tmp/

OmniParser/omnitool/gradio/agent/anthropic_agent.py

@@ -0,0 +1,162 @@
+"""
+Agentic sampling loop that calls the Anthropic API and local implenmentation of anthropic-defined computer use tools.
+"""
+import asyncio
+import platform
+from collections.abc import Callable
+from datetime import datetime
+from enum import StrEnum
+from typing import Any, cast
+
+from anthropic import Anthropic, AnthropicBedrock, AnthropicVertex, APIResponse
+from anthropic.types import (
+    ToolResultBlockParam,
+)
+from anthropic.types.beta import (
+    BetaContentBlock,
+    BetaContentBlockParam,
+    BetaImageBlockParam,
+    BetaMessage,
+    BetaMessageParam,
+    BetaTextBlockParam,
+    BetaToolResultBlockParam,
+)
+from anthropic.types import TextBlock
+from anthropic.types.beta import BetaMessage, BetaTextBlock, BetaToolUseBlock
+
+from tools import ComputerTool, ToolCollection, ToolResult
+
+from PIL import Image
+from io import BytesIO
+import gradio as gr
+from typing import Dict
+
+BETA_FLAG = "computer-use-2024-10-22"
+
+class APIProvider(StrEnum):
+    ANTHROPIC = "anthropic"
+    BEDROCK = "bedrock"
+    VERTEX = "vertex"
+
+SYSTEM_PROMPT = f"""<SYSTEM_CAPABILITY>
+* You are utilizing a Windows system with internet access.
+* The current date is {datetime.today().strftime('%A, %B %d, %Y')}.
+</SYSTEM_CAPABILITY>
+"""
+
+class AnthropicActor:
+    def __init__(
+        self, 
+        model: str, 
+        provider: APIProvider,
+        api_key: str,
+        api_response_callback: Callable[[APIResponse[BetaMessage]], None],
+        max_tokens: int = 4096,
+        only_n_most_recent_images: int | None = None,
+        print_usage: bool = True,
+    ):
+        self.model = model
+        self.provider = provider
+        self.api_key = api_key
+        self.api_response_callback = api_response_callback
+        self.max_tokens = max_tokens
+        self.only_n_most_recent_images = only_n_most_recent_images
+        
+        self.tool_collection = ToolCollection(ComputerTool())
+
+        self.system = SYSTEM_PROMPT
+        
+        self.total_token_usage = 0
+        self.total_cost = 0
+        self.print_usage = print_usage
+
+        # Instantiate the appropriate API client based on the provider
+        if provider == APIProvider.ANTHROPIC:
+            self.client = Anthropic(api_key=api_key)
+        elif provider == APIProvider.VERTEX:
+            self.client = AnthropicVertex()
+        elif provider == APIProvider.BEDROCK:
+            self.client = AnthropicBedrock()
+
+    def __call__(
+        self, 
+        *,
+        messages: list[BetaMessageParam]
+    ):
+        """
+        Generate a response given history messages.
+        """
+        if self.only_n_most_recent_images:
+            _maybe_filter_to_n_most_recent_images(messages, self.only_n_most_recent_images)
+
+        # Call the API synchronously
+        raw_response = self.client.beta.messages.with_raw_response.create(
+            max_tokens=self.max_tokens,
+            messages=messages,
+            model=self.model,
+            system=self.system,
+            tools=self.tool_collection.to_params(),
+            betas=["computer-use-2024-10-22"],
+        )
+
+        self.api_response_callback(cast(APIResponse[BetaMessage], raw_response))
+
+        response = raw_response.parse()
+        print(f"AnthropicActor response: {response}")
+
+        self.total_token_usage += response.usage.input_tokens + response.usage.output_tokens
+        self.total_cost += (response.usage.input_tokens * 3 / 1000000 + response.usage.output_tokens * 15 / 1000000)
+        
+        if self.print_usage:
+            print(f"Claude total token usage so far: {self.total_token_usage}, total cost so far: $USD{self.total_cost}")
+        
+        return response
+
+
+def _maybe_filter_to_n_most_recent_images(
+    messages: list[BetaMessageParam],
+    images_to_keep: int,
+    min_removal_threshold: int = 10,
+):
+    """
+    With the assumption that images are screenshots that are of diminishing value as
+    the conversation progresses, remove all but the final `images_to_keep` tool_result
+    images in place, with a chunk of min_removal_threshold to reduce the amount we
+    break the implicit prompt cache.
+    """
+    if images_to_keep is None:
+        return messages
+
+    tool_result_blocks = cast(
+        list[ToolResultBlockParam],
+        [
+            item
+            for message in messages
+            for item in (
+                message["content"] if isinstance(message["content"], list) else []
+            )
+            if isinstance(item, dict) and item.get("type") == "tool_result"
+        ],
+    )
+
+    total_images = sum(
+        1
+        for tool_result in tool_result_blocks
+        for content in tool_result.get("content", [])
+        if isinstance(content, dict) and content.get("type") == "image"
+    )
+
+    images_to_remove = total_images - images_to_keep
+    # for better cache behavior, we want to remove in chunks
+    images_to_remove -= images_to_remove % min_removal_threshold
+
+    for tool_result in tool_result_blocks:
+        if isinstance(tool_result.get("content"), list):
+            new_content = []
+            for content in tool_result.get("content", []):
+                if isinstance(content, dict) and content.get("type") == "image":
+                    if images_to_remove > 0:
+                        images_to_remove -= 1
+                        continue
+                new_content.append(content)
+            tool_result["content"] = new_content

OmniParser/omnitool/gradio/agent/llm_utils/groqclient.py

@@ -0,0 +1,59 @@
+from groq import Groq
+import os
+from .utils import is_image_path
+
+def run_groq_interleaved(messages: list, system: str, model_name: str, api_key: str, max_tokens=256, temperature=0.6):
+    """
+    Run a chat completion through Groq's API, ignoring any images in the messages.
+    """
+    api_key = api_key or os.environ.get("GROQ_API_KEY")
+    if not api_key:
+        raise ValueError("GROQ_API_KEY is not set")
+    
+    client = Groq(api_key=api_key)
+    # avoid using system messages for R1
+    final_messages = [{"role": "user", "content": system}]
+
+    if isinstance(messages, list):
+        for item in messages:
+            if isinstance(item, dict):
+                # For dict items, concatenate all text content, ignoring images
+                text_contents = []
+                for cnt in item["content"]:
+                    if isinstance(cnt, str):
+                        if not is_image_path(cnt):  # Skip image paths
+                            text_contents.append(cnt)
+                    else:
+                        text_contents.append(str(cnt))
+                
+                if text_contents:  # Only add if there's text content
+                    message = {"role": "user", "content": " ".join(text_contents)}
+                    final_messages.append(message)
+            else:  # str
+                message = {"role": "user", "content": item}
+                final_messages.append(message)
+    
+    elif isinstance(messages, str):
+        final_messages.append({"role": "user", "content": messages})
+
+    try:
+        completion = client.chat.completions.create(
+            model="deepseek-r1-distill-llama-70b",
+            messages=final_messages,
+            temperature=0.6,
+            max_completion_tokens=max_tokens,
+            top_p=0.95,
+            stream=False,
+            reasoning_format="raw"
+        )
+        
+        response = completion.choices[0].message.content
+        final_answer = response.split('</think>\n')[-1] if '</think>' in response else response
+        final_answer = final_answer.replace("<output>", "").replace("</output>", "")
+        token_usage = completion.usage.total_tokens
+        
+        return final_answer, token_usage
+    except Exception as e:
+        print(f"Error in interleaved Groq: {e}")
+
+        return str(e), 0

OmniParser/omnitool/gradio/agent/llm_utils/oaiclient.py

@@ -0,0 +1,62 @@
+import os
+import logging
+import base64
+import requests
+from .utils import is_image_path, encode_image
+
+def run_oai_interleaved(messages: list, system: str, model_name: str, api_key: str, max_tokens=256, temperature=0, provider_base_url: str = "https://api.openai.com/v1"):    
+    headers = {"Content-Type": "application/json",
+               "Authorization": f"Bearer {api_key}"}
+    final_messages = [{"role": "system", "content": system}]
+
+    if type(messages) == list:
+        for item in messages:
+            contents = []
+            if isinstance(item, dict):
+                for cnt in item["content"]:
+                    if isinstance(cnt, str):
+                        if is_image_path(cnt) and 'o3-mini' not in model_name:
+                            # 03 mini does not support images
+                            base64_image = encode_image(cnt)
+                            content = {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{base64_image}"}}
+                        else:
+                            content = {"type": "text", "text": cnt}
+                    else:
+                        # in this case it is a text block from anthropic
+                        content = {"type": "text", "text": str(cnt)}
+                        
+                    contents.append(content)
+                    
+                message = {"role": 'user', "content": contents}
+            else:  # str
+                contents.append({"type": "text", "text": item})
+                message = {"role": "user", "content": contents}
+            
+            final_messages.append(message)
+
+    
+    elif isinstance(messages, str):
+        final_messages = [{"role": "user", "content": messages}]
+
+    payload = {
+        "model": model_name,
+        "messages": final_messages,
+    }
+    if 'o1' in model_name or 'o3-mini' in model_name:
+        payload['reasoning_effort'] = 'low'
+        payload['max_completion_tokens'] = max_tokens
+    else:
+        payload['max_tokens'] = max_tokens
+
+    response = requests.post(
+        f"{provider_base_url}/chat/completions", headers=headers, json=payload
+    )
+
+
+    try:
+        text = response.json()['choices'][0]['message']['content']
+        token_usage = int(response.json()['usage']['total_tokens'])
+        return text, token_usage
+    except Exception as e:
+        print(f"Error in interleaved openAI: {e}. This may due to your invalid API key. Please check the response: {response.json()} ")
+        return response.json()

OmniParser/omnitool/gradio/agent/llm_utils/omniparserclient.py

@@ -0,0 +1,44 @@
+import requests
+import base64
+from pathlib import Path
+from tools.screen_capture import get_screenshot
+from agent.llm_utils.utils import encode_image
+
+OUTPUT_DIR = "./tmp/outputs"
+
+class OmniParserClient:
+    def __init__(self, 
+                 url: str) -> None:
+        self.url = url
+
+    def __call__(self,):
+        screenshot, screenshot_path = get_screenshot()
+        screenshot_path = str(screenshot_path)
+        image_base64 = encode_image(screenshot_path)
+        response = requests.post(self.url, json={"base64_image": image_base64})
+        response_json = response.json()
+        print('omniparser latency:', response_json['latency'])
+
+        som_image_data = base64.b64decode(response_json['som_image_base64'])
+        screenshot_path_uuid = Path(screenshot_path).stem.replace("screenshot_", "")
+        som_screenshot_path = f"{OUTPUT_DIR}/screenshot_som_{screenshot_path_uuid}.png"
+        with open(som_screenshot_path, "wb") as f:
+            f.write(som_image_data)
+        
+        response_json['width'] = screenshot.size[0]
+        response_json['height'] = screenshot.size[1]
+        response_json['original_screenshot_base64'] = image_base64
+        response_json['screenshot_uuid'] = screenshot_path_uuid
+        response_json = self.reformat_messages(response_json)
+        return response_json
+    
+    def reformat_messages(self, response_json: dict):
+        screen_info = ""
+        for idx, element in enumerate(response_json["parsed_content_list"]):
+            element['idx'] = idx
+            if element['type'] == 'text':
+                screen_info += f'ID: {idx}, Text: {element["content"]}\n'
+            elif element['type'] == 'icon':
+                screen_info += f'ID: {idx}, Icon: {element["content"]}\n'
+        response_json['screen_info'] = screen_info
+        return response_json

OmniParser/omnitool/gradio/agent/llm_utils/utils.py

@@ -0,0 +1,13 @@
+import base64
+
+def is_image_path(text):
+    image_extensions = (".jpg", ".jpeg", ".png", ".gif", ".bmp", ".tiff", ".tif")
+    if text.endswith(image_extensions):
+        return True
+    else:
+        return False
+
+def encode_image(image_path):
+    """Encode image file to base64."""
+    with open(image_path, "rb") as image_file:
+        return base64.b64encode(image_file.read()).decode("utf-8")

OmniParser/omnitool/gradio/agent/vlm_agent.py

@@ -0,0 +1,353 @@
+import json
+from collections.abc import Callable
+from typing import cast, Callable
+import uuid
+from PIL import Image, ImageDraw
+import base64
+from io import BytesIO
+
+from anthropic import APIResponse
+from anthropic.types import ToolResultBlockParam
+from anthropic.types.beta import BetaMessage, BetaTextBlock, BetaToolUseBlock, BetaMessageParam, BetaUsage
+
+from agent.llm_utils.oaiclient import run_oai_interleaved
+from agent.llm_utils.groqclient import run_groq_interleaved
+from agent.llm_utils.utils import is_image_path
+import time
+import re
+
+OUTPUT_DIR = "./tmp/outputs"
+
+def extract_data(input_string, data_type):
+    # Regular expression to extract content starting from '```python' until the end if there are no closing backticks
+    pattern = f"```{data_type}" + r"(.*?)(```|$)"
+    # Extract content
+    # re.DOTALL allows '.' to match newlines as well
+    matches = re.findall(pattern, input_string, re.DOTALL)
+    # Return the first match if exists, trimming whitespace and ignoring potential closing backticks
+    return matches[0][0].strip() if matches else input_string
+
+class VLMAgent:
+    def __init__(
+        self,
+        model: str, 
+        provider: str, 
+        api_key: str,
+        output_callback: Callable, 
+        api_response_callback: Callable,
+        max_tokens: int = 4096,
+        only_n_most_recent_images: int | None = None,
+        print_usage: bool = True,
+    ):
+        if model == "omniparser + gpt-4o":
+            self.model = "gpt-4o-2024-11-20"
+        elif model == "omniparser + R1":
+            self.model = "deepseek-r1-distill-llama-70b"
+        elif model == "omniparser + qwen2.5vl":
+            self.model = "qwen2.5-vl-72b-instruct"
+        elif model == "omniparser + o1":
+            self.model = "o1"
+        elif model == "omniparser + o3-mini":
+            self.model = "o3-mini"
+        else:
+            raise ValueError(f"Model {model} not supported")
+        
+
+        self.provider = provider
+        self.api_key = api_key
+        self.api_response_callback = api_response_callback
+        self.max_tokens = max_tokens
+        self.only_n_most_recent_images = only_n_most_recent_images
+        self.output_callback = output_callback
+
+        self.print_usage = print_usage
+        self.total_token_usage = 0
+        self.total_cost = 0
+        self.step_count = 0
+
+        self.system = ''
+           
+    def __call__(self, messages: list, parsed_screen: list[str, list, dict]):
+        self.step_count += 1
+        image_base64 = parsed_screen['original_screenshot_base64']
+        latency_omniparser = parsed_screen['latency']
+        self.output_callback(f'-- Step {self.step_count}: --', sender="bot")
+        screen_info = str(parsed_screen['screen_info'])
+        screenshot_uuid = parsed_screen['screenshot_uuid']
+        screen_width, screen_height = parsed_screen['width'], parsed_screen['height']
+
+        boxids_and_labels = parsed_screen["screen_info"]
+        system = self._get_system_prompt(boxids_and_labels)
+
+        # drop looping actions msg, byte image etc
+        planner_messages = messages
+        _remove_som_images(planner_messages)
+        _maybe_filter_to_n_most_recent_images(planner_messages, self.only_n_most_recent_images)
+
+        if isinstance(planner_messages[-1], dict):
+            if not isinstance(planner_messages[-1]["content"], list):
+                planner_messages[-1]["content"] = [planner_messages[-1]["content"]]
+            planner_messages[-1]["content"].append(f"{OUTPUT_DIR}/screenshot_{screenshot_uuid}.png")
+            planner_messages[-1]["content"].append(f"{OUTPUT_DIR}/screenshot_som_{screenshot_uuid}.png")
+
+        start = time.time()
+        if "gpt" in self.model or "o1" in self.model or "o3-mini" in self.model:
+            vlm_response, token_usage = run_oai_interleaved(
+                messages=planner_messages,
+                system=system,
+                model_name=self.model,
+                api_key=self.api_key,
+                max_tokens=self.max_tokens,
+                provider_base_url="https://api.openai.com/v1",
+                temperature=0,
+            )
+            print(f"oai token usage: {token_usage}")
+            self.total_token_usage += token_usage
+            if 'gpt' in self.model:
+                self.total_cost += (token_usage * 2.5 / 1000000)  # https://openai.com/api/pricing/
+            elif 'o1' in self.model:
+                self.total_cost += (token_usage * 15 / 1000000)  # https://openai.com/api/pricing/
+            elif 'o3-mini' in self.model:
+                self.total_cost += (token_usage * 1.1 / 1000000)  # https://openai.com/api/pricing/
+        elif "r1" in self.model:
+            vlm_response, token_usage = run_groq_interleaved(
+                messages=planner_messages,
+                system=system,
+                model_name=self.model,
+                api_key=self.api_key,
+                max_tokens=self.max_tokens,
+            )
+            print(f"groq token usage: {token_usage}")
+            self.total_token_usage += token_usage
+            self.total_cost += (token_usage * 0.99 / 1000000)
+        elif "qwen" in self.model:
+            vlm_response, token_usage = run_oai_interleaved(
+                messages=planner_messages,
+                system=system,
+                model_name=self.model,
+                api_key=self.api_key,
+                max_tokens=min(2048, self.max_tokens),
+                provider_base_url="https://dashscope.aliyuncs.com/compatible-mode/v1",
+                temperature=0,
+            )
+            print(f"qwen token usage: {token_usage}")
+            self.total_token_usage += token_usage
+            self.total_cost += (token_usage * 2.2 / 1000000)  # https://help.aliyun.com/zh/model-studio/getting-started/models?spm=a2c4g.11186623.0.0.74b04823CGnPv7#fe96cfb1a422a
+        else:
+            raise ValueError(f"Model {self.model} not supported")
+        latency_vlm = time.time() - start
+        self.output_callback(f"LLM: {latency_vlm:.2f}s, OmniParser: {latency_omniparser:.2f}s", sender="bot")
+
+        print(f"{vlm_response}")
+        
+        if self.print_usage:
+            print(f"Total token so far: {self.total_token_usage}. Total cost so far: $USD{self.total_cost:.5f}")
+        
+        vlm_response_json = extract_data(vlm_response, "json")
+        vlm_response_json = json.loads(vlm_response_json)
+
+        img_to_show_base64 = parsed_screen["som_image_base64"]
+        if "Box ID" in vlm_response_json:
+            try:
+                bbox = parsed_screen["parsed_content_list"][int(vlm_response_json["Box ID"])]["bbox"]
+                vlm_response_json["box_centroid_coordinate"] = [int((bbox[0] + bbox[2]) / 2 * screen_width), int((bbox[1] + bbox[3]) / 2 * screen_height)]
+                img_to_show_data = base64.b64decode(img_to_show_base64)
+                img_to_show = Image.open(BytesIO(img_to_show_data))
+
+                draw = ImageDraw.Draw(img_to_show)
+                x, y = vlm_response_json["box_centroid_coordinate"] 
+                radius = 10
+                draw.ellipse((x - radius, y - radius, x + radius, y + radius), fill='red')
+                draw.ellipse((x - radius*3, y - radius*3, x + radius*3, y + radius*3), fill=None, outline='red', width=2)
+
+                buffered = BytesIO()
+                img_to_show.save(buffered, format="PNG")
+                img_to_show_base64 = base64.b64encode(buffered.getvalue()).decode("utf-8")
+            except:
+                print(f"Error parsing: {vlm_response_json}")
+                pass
+        self.output_callback(f'<img src="data:image/png;base64,{img_to_show_base64}">', sender="bot")
+        self.output_callback(
+                    f'<details>'
+                    f'  <summary>Parsed Screen elemetns by OmniParser</summary>'
+                    f'  <pre>{screen_info}</pre>'
+                    f'</details>',
+                    sender="bot"
+                )
+        vlm_plan_str = ""
+        for key, value in vlm_response_json.items():
+            if key == "Reasoning":
+                vlm_plan_str += f'{value}'
+            else:
+                vlm_plan_str += f'\n{key}: {value}'
+
+        # construct the response so that anthropicExcutor can execute the tool
+        response_content = [BetaTextBlock(text=vlm_plan_str, type='text')]
+        if 'box_centroid_coordinate' in vlm_response_json:
+            move_cursor_block = BetaToolUseBlock(id=f'toolu_{uuid.uuid4()}',
+                                            input={'action': 'mouse_move', 'coordinate': vlm_response_json["box_centroid_coordinate"]},
+                                            name='computer', type='tool_use')
+            response_content.append(move_cursor_block)
+
+        if vlm_response_json["Next Action"] == "None":
+            print("Task paused/completed.")
+        elif vlm_response_json["Next Action"] == "type":
+            sim_content_block = BetaToolUseBlock(id=f'toolu_{uuid.uuid4()}',
+                                        input={'action': vlm_response_json["Next Action"], 'text': vlm_response_json["value"]},
+                                        name='computer', type='tool_use')
+            response_content.append(sim_content_block)
+        else:
+            sim_content_block = BetaToolUseBlock(id=f'toolu_{uuid.uuid4()}',
+                                            input={'action': vlm_response_json["Next Action"]},
+                                            name='computer', type='tool_use')
+            response_content.append(sim_content_block)
+        response_message = BetaMessage(id=f'toolu_{uuid.uuid4()}', content=response_content, model='', role='assistant', type='message', stop_reason='tool_use', usage=BetaUsage(input_tokens=0, output_tokens=0))
+        return response_message, vlm_response_json
+
+    def _api_response_callback(self, response: APIResponse):
+        self.api_response_callback(response)
+
+    def _get_system_prompt(self, screen_info: str = ""):
+        main_section = f"""
+You are using a Windows device.
+You are able to use a mouse and keyboard to interact with the computer based on the given task and screenshot.
+You can only interact with the desktop GUI (no terminal or application menu access).
+
+You may be given some history plan and actions, this is the response from the previous loop.
+You should carefully consider your plan base on the task, screenshot, and history actions.
+
+Here is the list of all detected bounding boxes by IDs on the screen and their description:{screen_info}
+
+Your available "Next Action" only include:
+- type: types a string of text.
+- left_click: move mouse to box id and left clicks.
+- right_click: move mouse to box id and right clicks.
+- double_click: move mouse to box id and double clicks.
+- hover: move mouse to box id.
+- scroll_up: scrolls the screen up to view previous content.
+- scroll_down: scrolls the screen down, when the desired button is not visible, or you need to see more content. 
+- wait: waits for 1 second for the device to load or respond.
+
+Based on the visual information from the screenshot image and the detected bounding boxes, please determine the next action, the Box ID you should operate on (if action is one of 'type', 'hover', 'scroll_up', 'scroll_down', 'wait', there should be no Box ID field), and the value (if the action is 'type') in order to complete the task.
+
+Output format:
+```json
+{{
+    "Reasoning": str, # describe what is in the current screen, taking into account the history, then describe your step-by-step thoughts on how to achieve the task, choose one action from available actions at a time.
+    "Next Action": "action_type, action description" | "None" # one action at a time, describe it in short and precisely. 
+    "Box ID": n,
+    "value": "xxx" # only provide value field if the action is type, else don't include value key
+}}
+```
+
+One Example:
+```json
+{{  
+    "Reasoning": "The current screen shows google result of amazon, in previous action I have searched amazon on google. Then I need to click on the first search results to go to amazon.com.",
+    "Next Action": "left_click",
+    "Box ID": m
+}}
+```
+
+Another Example:
+```json
+{{
+    "Reasoning": "The current screen shows the front page of amazon. There is no previous action. Therefore I need to type "Apple watch" in the search bar.",
+    "Next Action": "type",
+    "Box ID": n,
+    "value": "Apple watch"
+}}
+```
+
+Another Example:
+```json
+{{
+    "Reasoning": "The current screen does not show 'submit' button, I need to scroll down to see if the button is available.",
+    "Next Action": "scroll_down",
+}}
+```
+
+IMPORTANT NOTES:
+1. You should only give a single action at a time.
+
+"""
+        thinking_model = "r1" in self.model
+        if not thinking_model:
+            main_section += """
+2. You should give an analysis to the current screen, and reflect on what has been done by looking at the history, then describe your step-by-step thoughts on how to achieve the task.
+
+"""
+        else:
+            main_section += """
+2. In <think> XML tags give an analysis to the current screen, and reflect on what has been done by looking at the history, then describe your step-by-step thoughts on how to achieve the task. In <output> XML tags put the next action prediction JSON.
+
+"""
+        main_section += """
+3. Attach the next action prediction in the "Next Action".
+4. You should not include other actions, such as keyboard shortcuts.
+5. When the task is completed, don't complete additional actions. You should say "Next Action": "None" in the json field.
+6. The tasks involve buying multiple products or navigating through multiple pages. You should break it into subgoals and complete each subgoal one by one in the order of the instructions.
+7. avoid choosing the same action/elements multiple times in a row, if it happens, reflect to yourself, what may have gone wrong, and predict a different action.
+8. If you are prompted with login information page or captcha page, or you think it need user's permission to do the next action, you should say "Next Action": "None" in the json field.
+""" 
+
+        return main_section
+
+def _remove_som_images(messages):
+    for msg in messages:
+        msg_content = msg["content"]
+        if isinstance(msg_content, list):
+            msg["content"] = [
+                cnt for cnt in msg_content 
+                if not (isinstance(cnt, str) and 'som' in cnt and is_image_path(cnt))
+            ]
+
+
+def _maybe_filter_to_n_most_recent_images(
+    messages: list[BetaMessageParam],
+    images_to_keep: int,
+    min_removal_threshold: int = 10,
+):
+    """
+    With the assumption that images are screenshots that are of diminishing value as
+    the conversation progresses, remove all but the final `images_to_keep` tool_result
+    images in place
+    """
+    if images_to_keep is None:
+        return messages
+
+    total_images = 0
+    for msg in messages:
+        for cnt in msg.get("content", []):
+            if isinstance(cnt, str) and is_image_path(cnt):
+                total_images += 1
+            elif isinstance(cnt, dict) and cnt.get("type") == "tool_result":
+                for content in cnt.get("content", []):
+                    if isinstance(content, dict) and content.get("type") == "image":
+                        total_images += 1
+
+    images_to_remove = total_images - images_to_keep
+    
+    for msg in messages:
+        msg_content = msg["content"]
+        if isinstance(msg_content, list):
+            new_content = []
+            for cnt in msg_content:
+                # Remove images from SOM or screenshot as needed
+                if isinstance(cnt, str) and is_image_path(cnt):
+                    if images_to_remove > 0:
+                        images_to_remove -= 1
+                        continue
+                # VLM shouldn't use anthropic screenshot tool so shouldn't have these but in case it does, remove as needed
+                elif isinstance(cnt, dict) and cnt.get("type") == "tool_result":
+                    new_tool_result_content = []
+                    for tool_result_entry in cnt.get("content", []):
+                        if isinstance(tool_result_entry, dict) and tool_result_entry.get("type") == "image":
+                            if images_to_remove > 0:
+                                images_to_remove -= 1
+                                continue
+                        new_tool_result_content.append(tool_result_entry)
+                    cnt["content"] = new_tool_result_content
+                # Append fixed content to current message's content list
+                new_content.append(cnt)
+            msg["content"] = new_content

OmniParser/omnitool/gradio/agent/vlm_agent_with_orchestrator.py

@@ -0,0 +1,498 @@
+import json
+from collections.abc import Callable
+from typing import cast, Callable
+import uuid
+from PIL import Image, ImageDraw
+import base64
+from io import BytesIO
+import copy
+from pathlib import Path
+from datetime import datetime
+from anthropic import APIResponse
+from anthropic.types import ToolResultBlockParam
+from anthropic.types.beta import BetaMessage, BetaTextBlock, BetaToolUseBlock, BetaMessageParam, BetaUsage
+
+from agent.llm_utils.oaiclient import run_oai_interleaved
+from agent.llm_utils.groqclient import run_groq_interleaved
+from agent.llm_utils.utils import is_image_path
+import time
+import re
+import os
+OUTPUT_DIR = "./tmp/outputs"
+ORCHESTRATOR_LEDGER_PROMPT = """
+Recall we are working on the following request:
+
+{task}
+
+To make progress on the request, please answer the following questions, including necessary reasoning:
+
+    - Is the request fully satisfied? (True if complete, or False if the original request has yet to be SUCCESSFULLY and FULLY addressed)
+    - Are we in a loop where we are repeating the same requests and / or getting the same responses as before? Loops can span multiple turns, and can include repeated actions like scrolling up or down more than a handful of times.
+    - Are we making forward progress? (True if just starting, or recent messages are adding value. False if recent messages show evidence of being stuck in a loop or if there is evidence of significant barriers to success such as the inability to read from a required file)
+    - What instruction or question would you give in order to complete the task? 
+
+Please output an answer in pure JSON format according to the following schema. The JSON object must be parsable as-is. DO NOT OUTPUT ANYTHING OTHER THAN JSON, AND DO NOT DEVIATE FROM THIS SCHEMA:
+
+    {{
+       "is_request_satisfied": {{
+            "reason": string,
+            "answer": boolean
+        }},
+        "is_in_loop": {{
+            "reason": string,
+            "answer": boolean
+        }},
+        "is_progress_being_made": {{
+            "reason": string,
+            "answer": boolean
+        }},
+        "instruction_or_question": {{
+            "reason": string,
+            "answer": string
+        }}
+    }}
+"""
+
+def extract_data(input_string, data_type):
+    # Regular expression to extract content starting from '```python' until the end if there are no closing backticks
+    pattern = f"```{data_type}" + r"(.*?)(```|$)"
+    # Extract content
+    # re.DOTALL allows '.' to match newlines as well
+    matches = re.findall(pattern, input_string, re.DOTALL)
+    # Return the first match if exists, trimming whitespace and ignoring potential closing backticks
+    return matches[0][0].strip() if matches else input_string
+
+class VLMOrchestratedAgent:
+    def __init__(
+        self,
+        model: str, 
+        provider: str, 
+        api_key: str,
+        output_callback: Callable, 
+        api_response_callback: Callable,
+        max_tokens: int = 4096,
+        only_n_most_recent_images: int | None = None,
+        print_usage: bool = True,
+        save_folder: str = None,
+    ):
+        if model == "omniparser + gpt-4o" or model == "omniparser + gpt-4o-orchestrated":
+            self.model = "gpt-4o-2024-11-20"
+        elif model == "omniparser + R1" or model == "omniparser + R1-orchestrated":
+            self.model = "deepseek-r1-distill-llama-70b"
+        elif model == "omniparser + qwen2.5vl" or model == "omniparser + qwen2.5vl-orchestrated":
+            self.model = "qwen2.5-vl-72b-instruct"
+        elif model == "omniparser + o1" or model == "omniparser + o1-orchestrated":
+            self.model = "o1"
+        elif model == "omniparser + o3-mini" or model == "omniparser + o3-mini-orchestrated":
+            self.model = "o3-mini"
+        else:
+            raise ValueError(f"Model {model} not supported")
+        
+
+        self.provider = provider
+        self.api_key = api_key
+        self.api_response_callback = api_response_callback
+        self.max_tokens = max_tokens
+        self.only_n_most_recent_images = only_n_most_recent_images
+        self.output_callback = output_callback
+        self.save_folder = save_folder
+        
+        self.print_usage = print_usage
+        self.total_token_usage = 0
+        self.total_cost = 0
+        self.step_count = 0
+        self.plan, self.ledger = None, None
+
+        self.system = ''
+    
+    def __call__(self, messages: list, parsed_screen: list[str, list, dict]):
+        if self.step_count == 0:
+            plan = self._initialize_task(messages)
+            self.output_callback(f'-- Plan: {plan} --', )
+            # update messages with the plan
+            messages.append({"role": "assistant", "content": plan})
+        else:
+            updated_ledger = self._update_ledger(messages)
+            self.output_callback(
+                f'<details>'
+                f'  <summary><strong>Task Progress Ledger (click to expand)</strong></summary>'
+                f'  <div style="padding: 10px; background-color: #f8f9fa; border-radius: 5px; margin-top: 5px;">'
+                f'    <pre>{updated_ledger}</pre>'
+                f'  </div>'
+                f'</details>',
+            )
+            # update messages with the ledger
+            messages.append({"role": "assistant", "content": updated_ledger})
+            self.ledger = updated_ledger
+
+        self.step_count += 1
+        # save the image to the output folder
+        with open(f"{self.save_folder}/screenshot_{self.step_count}.png", "wb") as f:
+            f.write(base64.b64decode(parsed_screen['original_screenshot_base64']))
+        with open(f"{self.save_folder}/som_screenshot_{self.step_count}.png", "wb") as f:
+            f.write(base64.b64decode(parsed_screen['som_image_base64']))
+
+        latency_omniparser = parsed_screen['latency']
+        screen_info = str(parsed_screen['screen_info'])
+        screenshot_uuid = parsed_screen['screenshot_uuid']
+        screen_width, screen_height = parsed_screen['width'], parsed_screen['height']
+
+        boxids_and_labels = parsed_screen["screen_info"]
+        system = self._get_system_prompt(boxids_and_labels)
+
+        # drop looping actions msg, byte image etc
+        planner_messages = messages
+        _remove_som_images(planner_messages)
+        _maybe_filter_to_n_most_recent_images(planner_messages, self.only_n_most_recent_images)
+
+        if isinstance(planner_messages[-1], dict):
+            if not isinstance(planner_messages[-1]["content"], list):
+                planner_messages[-1]["content"] = [planner_messages[-1]["content"]]
+            planner_messages[-1]["content"].append(f"{OUTPUT_DIR}/screenshot_{screenshot_uuid}.png")
+            planner_messages[-1]["content"].append(f"{OUTPUT_DIR}/screenshot_som_{screenshot_uuid}.png")
+
+        start = time.time()
+        if "gpt" in self.model or "o1" in self.model or "o3-mini" in self.model:
+            vlm_response, token_usage = run_oai_interleaved(
+                messages=planner_messages,
+                system=system,
+                model_name=self.model,
+                api_key=self.api_key,
+                max_tokens=self.max_tokens,
+                provider_base_url="https://api.openai.com/v1",
+                temperature=0,
+            )
+            print(f"oai token usage: {token_usage}")
+            self.total_token_usage += token_usage
+            if 'gpt' in self.model:
+                self.total_cost += (token_usage * 2.5 / 1000000)  # https://openai.com/api/pricing/
+            elif 'o1' in self.model:
+                self.total_cost += (token_usage * 15 / 1000000)  # https://openai.com/api/pricing/
+            elif 'o3-mini' in self.model:
+                self.total_cost += (token_usage * 1.1 / 1000000)  # https://openai.com/api/pricing/
+        elif "r1" in self.model:
+            vlm_response, token_usage = run_groq_interleaved(
+                messages=planner_messages,
+                system=system,
+                model_name=self.model,
+                api_key=self.api_key,
+                max_tokens=self.max_tokens,
+            )
+            print(f"groq token usage: {token_usage}")
+            self.total_token_usage += token_usage
+            self.total_cost += (token_usage * 0.99 / 1000000)
+        elif "qwen" in self.model:
+            vlm_response, token_usage = run_oai_interleaved(
+                messages=planner_messages,
+                system=system,
+                model_name=self.model,
+                api_key=self.api_key,
+                max_tokens=min(2048, self.max_tokens),
+                provider_base_url="https://dashscope.aliyuncs.com/compatible-mode/v1",
+                temperature=0,
+            )
+            print(f"qwen token usage: {token_usage}")
+            self.total_token_usage += token_usage
+            self.total_cost += (token_usage * 2.2 / 1000000)  # https://help.aliyun.com/zh/model-studio/getting-started/models?spm=a2c4g.11186623.0.0.74b04823CGnPv7#fe96cfb1a422a
+        else:
+            raise ValueError(f"Model {self.model} not supported")
+        latency_vlm = time.time() - start
+        
+        # Update step counter with both latencies
+        self.output_callback(f'<i>Step {self.step_count} | OmniParser: {latency_omniparser:.2f}s | LLM: {latency_vlm:.2f}s</i>', )
+
+        print(f"{vlm_response}")
+        
+        if self.print_usage:
+            print(f"Total token so far: {self.total_token_usage}. Total cost so far: $USD{self.total_cost:.5f}")
+        
+        vlm_response_json = extract_data(vlm_response, "json")
+        vlm_response_json = json.loads(vlm_response_json)
+
+        img_to_show_base64 = parsed_screen["som_image_base64"]
+        if "Box ID" in vlm_response_json:
+            try:
+                bbox = parsed_screen["parsed_content_list"][int(vlm_response_json["Box ID"])]["bbox"]
+                vlm_response_json["box_centroid_coordinate"] = [int((bbox[0] + bbox[2]) / 2 * screen_width), int((bbox[1] + bbox[3]) / 2 * screen_height)]
+                img_to_show_data = base64.b64decode(img_to_show_base64)
+                img_to_show = Image.open(BytesIO(img_to_show_data))
+
+                draw = ImageDraw.Draw(img_to_show)
+                x, y = vlm_response_json["box_centroid_coordinate"] 
+                radius = 10
+                draw.ellipse((x - radius, y - radius, x + radius, y + radius), fill='red')
+                draw.ellipse((x - radius*3, y - radius*3, x + radius*3, y + radius*3), fill=None, outline='red', width=2)
+
+                buffered = BytesIO()
+                img_to_show.save(buffered, format="PNG")
+                img_to_show_base64 = base64.b64encode(buffered.getvalue()).decode("utf-8")
+            except:
+                print(f"Error parsing: {vlm_response_json}")
+                pass
+        self.output_callback(f'<img src="data:image/png;base64,{img_to_show_base64}">', )
+        
+        # Display screen info in a collapsible dropdown
+        self.output_callback(
+            f'<details>'
+            f'  <summary><strong>Parsed Screen Elements (click to expand)</strong></summary>'
+            f'  <div style="padding: 10px; background-color: #f8f9fa; border-radius: 5px; margin-top: 5px;">'
+            f'    <pre>{screen_info}</pre>'
+            f'  </div>'
+            f'</details>',
+        )
+        
+        vlm_plan_str = ""
+        for key, value in vlm_response_json.items():
+            if key == "Reasoning":
+                vlm_plan_str += f'{value}'
+            else:
+                vlm_plan_str += f'\n{key}: {value}'
+
+        # construct the response so that anthropicExcutor can execute the tool
+        response_content = [BetaTextBlock(text=vlm_plan_str, type='text')]
+        if 'box_centroid_coordinate' in vlm_response_json:
+            move_cursor_block = BetaToolUseBlock(id=f'toolu_{uuid.uuid4()}',
+                                            input={'action': 'mouse_move', 'coordinate': vlm_response_json["box_centroid_coordinate"]},
+                                            name='computer', type='tool_use')
+            response_content.append(move_cursor_block)
+
+        if vlm_response_json["Next Action"] == "None":
+            print("Task paused/completed.")
+        elif vlm_response_json["Next Action"] == "type":
+            sim_content_block = BetaToolUseBlock(id=f'toolu_{uuid.uuid4()}',
+                                        input={'action': vlm_response_json["Next Action"], 'text': vlm_response_json["value"]},
+                                        name='computer', type='tool_use')
+            response_content.append(sim_content_block)
+        else:
+            sim_content_block = BetaToolUseBlock(id=f'toolu_{uuid.uuid4()}',
+                                            input={'action': vlm_response_json["Next Action"]},
+                                            name='computer', type='tool_use')
+            response_content.append(sim_content_block)
+        response_message = BetaMessage(id=f'toolu_{uuid.uuid4()}', content=response_content, model='', role='assistant', type='message', stop_reason='tool_use', usage=BetaUsage(input_tokens=0, output_tokens=0))
+
+        # save the intermediate step trajectory to the save folder
+        step_trajectory = {
+            "screenshot_path": f"{self.save_folder}/screenshot_{self.step_count}.png",
+            "som_screenshot_path": f"{self.save_folder}/som_screenshot_{self.step_count}.png",
+            "screen_info": screen_info,
+            "latency_omniparser": latency_omniparser,
+            "latency_vlm": latency_vlm,
+            "vlm_response_json": vlm_response_json,
+            'ledger': self.ledger,
+        }
+        with open(f"{self.save_folder}/trajectory.json", "a") as f:
+            f.write(json.dumps(step_trajectory))
+            f.write("\n")
+
+        return response_message, vlm_response_json
+
+    def _api_response_callback(self, response: APIResponse):
+        self.api_response_callback(response)
+
+    def _get_system_prompt(self, screen_info: str = ""):
+        main_section = f"""
+You are using a Windows device.
+You are able to use a mouse and keyboard to interact with the computer based on the given task and screenshot.
+You can only interact with the desktop GUI (no terminal or application menu access).
+
+You may be given some history plan and actions, this is the response from the previous loop.
+You should carefully consider your plan base on the task, screenshot, and history actions.
+
+Here is the list of all detected bounding boxes by IDs on the screen and their description:{screen_info}
+
+Your available "Next Action" only include:
+- type: types a string of text.
+- left_click: move mouse to box id and left clicks.
+- right_click: move mouse to box id and right clicks.
+- double_click: move mouse to box id and double clicks.
+- hover: move mouse to box id.
+- scroll_up: scrolls the screen up to view previous content.
+- scroll_down: scrolls the screen down, when the desired button is not visible, or you need to see more content. 
+- wait: waits for 1 second for the device to load or respond.
+
+Based on the visual information from the screenshot image and the detected bounding boxes, please determine the next action, the Box ID you should operate on (if action is one of 'type', 'hover', 'scroll_up', 'scroll_down', 'wait', there should be no Box ID field), and the value (if the action is 'type') in order to complete the task.
+
+Output format:
+```json
+{{
+    "Reasoning": str, # describe what is in the current screen, taking into account the history, then describe your step-by-step thoughts on how to achieve the task, choose one action from available actions at a time.
+    "Next Action": "action_type, action description" | "None" # one action at a time, describe it in short and precisely. 
+    "Box ID": n,
+    "value": "xxx" # only provide value field if the action is type, else don't include value key
+}}
+```
+
+One Example:
+```json
+{{  
+    "Reasoning": "The current screen shows google result of amazon, in previous action I have searched amazon on google. Then I need to click on the first search results to go to amazon.com.",
+    "Next Action": "left_click",
+    "Box ID": m
+}}
+```
+
+Another Example:
+```json
+{{
+    "Reasoning": "The current screen shows the front page of amazon. There is no previous action. Therefore I need to type "Apple watch" in the search bar.",
+    "Next Action": "type",
+    "Box ID": n,
+    "value": "Apple watch"
+}}
+```
+
+Another Example:
+```json
+{{
+    "Reasoning": "The current screen does not show 'submit' button, I need to scroll down to see if the button is available.",
+    "Next Action": "scroll_down",
+}}
+```
+
+IMPORTANT NOTES:
+1. You should only give a single action at a time.
+
+"""
+        thinking_model = "r1" in self.model
+        if not thinking_model:
+            main_section += """
+2. You should give an analysis to the current screen, and reflect on what has been done by looking at the history, then describe your step-by-step thoughts on how to achieve the task.
+
+"""
+        else:
+            main_section += """
+2. In <think> XML tags give an analysis to the current screen, and reflect on what has been done by looking at the history, then describe your step-by-step thoughts on how to achieve the task. In <output> XML tags put the next action prediction JSON.
+
+"""
+        main_section += """
+3. Attach the next action prediction in the "Next Action".
+4. You should not include other actions, such as keyboard shortcuts.
+5. When the task is completed, don't complete additional actions. You should say "Next Action": "None" in the json field.
+6. The tasks involve buying multiple products or navigating through multiple pages. You should break it into subgoals and complete each subgoal one by one in the order of the instructions.
+7. avoid choosing the same action/elements multiple times in a row, if it happens, reflect to yourself, what may have gone wrong, and predict a different action.
+8. If you are prompted with login information page or captcha page, or you think it need user's permission to do the next action, you should say "Next Action": "None" in the json field.
+""" 
+
+        return main_section
+
+    def _initialize_task(self, messages: list):
+        self._task = messages[0]["content"]
+        # make a plan
+        plan_prompt = self._get_plan_prompt(self._task)
+        input_message = copy.deepcopy(messages)
+        input_message.append({"role": "user", "content": plan_prompt})
+        vlm_response, token_usage = run_oai_interleaved(
+                messages=input_message,
+                system="",
+                model_name=self.model,
+                api_key=self.api_key,
+                max_tokens=self.max_tokens,
+                provider_base_url="https://api.openai.com/v1",
+                temperature=0,
+            )
+        plan = extract_data(vlm_response, "json")
+        
+        # Create a filename with timestamp
+        plan_filename = f"plan.json"
+        plan_path = os.path.join(self.save_folder, plan_filename)
+        
+        # Save the plan to a file
+        try:
+            with open(plan_path, "w") as f:
+                f.write(plan)
+            print(f"Plan successfully saved to {plan_path}")
+        except Exception as e:
+            print(f"Error saving plan to {plan_path}: {str(e)}")
+        
+        return plan
+
+    def _update_ledger(self, messages):
+        # tobe implemented
+        # update the ledger with the current task and plan
+        # return the updated ledger
+        update_ledger_prompt = ORCHESTRATOR_LEDGER_PROMPT.format(task=self._task)
+        input_message = copy.deepcopy(messages)
+        input_message.append({"role": "user", "content": update_ledger_prompt})
+        vlm_response, token_usage = run_oai_interleaved(
+                messages=input_message,
+                system="",
+                model_name=self.model,
+                api_key=self.api_key,
+                max_tokens=self.max_tokens,
+                provider_base_url="https://api.openai.com/v1",
+                temperature=0,
+            )
+        updated_ledger = extract_data(vlm_response, "json")
+        return updated_ledger
+    
+    def _get_plan_prompt(self, task):
+        plan_prompt = f"""
+        please devise a short bullet-point plan for addressing the original user task: {task}
+        You should write your plan in a json dict, e.g:```json
+{{
+'step 1': xxx,
+'step 2': xxxx,
+...
+}}```
+        Now start your answer directly.
+        """
+        return plan_prompt
+
+def _remove_som_images(messages):
+    for msg in messages:
+        msg_content = msg["content"]
+        if isinstance(msg_content, list):
+            msg["content"] = [
+                cnt for cnt in msg_content 
+                if not (isinstance(cnt, str) and 'som' in cnt and is_image_path(cnt))
+            ]
+
+
+def _maybe_filter_to_n_most_recent_images(
+    messages: list[BetaMessageParam],
+    images_to_keep: int,
+    min_removal_threshold: int = 10,
+):
+    """
+    With the assumption that images are screenshots that are of diminishing value as
+    the conversation progresses, remove all but the final `images_to_keep` tool_result
+    images in place
+    """
+    if images_to_keep is None:
+        return messages
+
+    total_images = 0
+    for msg in messages:
+        for cnt in msg.get("content", []):
+            if isinstance(cnt, str) and is_image_path(cnt):
+                total_images += 1
+            elif isinstance(cnt, dict) and cnt.get("type") == "tool_result":
+                for content in cnt.get("content", []):
+                    if isinstance(content, dict) and content.get("type") == "image":
+                        total_images += 1
+
+    images_to_remove = total_images - images_to_keep
+    
+    for msg in messages:
+        msg_content = msg["content"]
+        if isinstance(msg_content, list):
+            new_content = []
+            for cnt in msg_content:
+                # Remove images from SOM or screenshot as needed
+                if isinstance(cnt, str) and is_image_path(cnt):
+                    if images_to_remove > 0:
+                        images_to_remove -= 1
+                        continue
+                # VLM shouldn't use anthropic screenshot tool so shouldn't have these but in case it does, remove as needed
+                elif isinstance(cnt, dict) and cnt.get("type") == "tool_result":
+                    new_tool_result_content = []
+                    for tool_result_entry in cnt.get("content", []):
+                        if isinstance(tool_result_entry, dict) and tool_result_entry.get("type") == "image":
+                            if images_to_remove > 0:
+                                images_to_remove -= 1
+                                continue
+                        new_tool_result_content.append(tool_result_entry)
+                    cnt["content"] = new_tool_result_content
+                # Append fixed content to current message's content list
+                new_content.append(cnt)
+            msg["content"] = new_content

OmniParser/omnitool/gradio/app.py

@@ -0,0 +1,426 @@
+"""
+python app.py --windows_host_url localhost:8006 --omniparser_server_url localhost:8000
+"""
+
+import os
+from datetime import datetime
+from enum import StrEnum
+from functools import partial
+from pathlib import Path
+from typing import cast
+import argparse
+import gradio as gr
+from anthropic import APIResponse
+from anthropic.types import TextBlock
+from anthropic.types.beta import BetaMessage, BetaTextBlock, BetaToolUseBlock
+from anthropic.types.tool_use_block import ToolUseBlock
+from loop import (
+    APIProvider,
+    sampling_loop_sync,
+)
+from tools import ToolResult
+import requests
+from requests.exceptions import RequestException
+import base64
+
+CONFIG_DIR = Path("~/.anthropic").expanduser()
+API_KEY_FILE = CONFIG_DIR / "api_key"
+
+INTRO_TEXT = '''
+OmniParser lets you turn any vision-langauge model into an AI agent. We currently support **OpenAI (4o/o1/o3-mini), DeepSeek (R1), Qwen (2.5VL) or Anthropic Computer Use (Sonnet).**
+
+Type a message and press submit to start OmniTool. Press stop to pause, and press the trash icon in the chat to clear the message history.
+'''
+
+def parse_arguments():
+
+    parser = argparse.ArgumentParser(description="Gradio App")
+    parser.add_argument("--windows_host_url", type=str, default='localhost:8006')
+    parser.add_argument("--omniparser_server_url", type=str, default="localhost:8000")
+    return parser.parse_args()
+args = parse_arguments()
+
+
+class Sender(StrEnum):
+    USER = "user"
+    BOT = "assistant"
+    TOOL = "tool"
+
+
+def setup_state(state):
+    if "messages" not in state:
+        state["messages"] = []
+    if "model" not in state:
+        state["model"] = "omniparser + gpt-4o"
+    if "provider" not in state:
+        state["provider"] = "openai"
+    if "openai_api_key" not in state:  # Fetch API keys from environment variables
+        state["openai_api_key"] = os.getenv("OPENAI_API_KEY", "")
+    if "anthropic_api_key" not in state:
+        state["anthropic_api_key"] = os.getenv("ANTHROPIC_API_KEY", "")
+    if "api_key" not in state:
+        state["api_key"] = ""
+    if "auth_validated" not in state:
+        state["auth_validated"] = False
+    if "responses" not in state:
+        state["responses"] = {}
+    if "tools" not in state:
+        state["tools"] = {}
+    if "only_n_most_recent_images" not in state:
+        state["only_n_most_recent_images"] = 2
+    if 'chatbot_messages' not in state:
+        state['chatbot_messages'] = []
+    if 'stop' not in state:
+        state['stop'] = False
+
+async def main(state):
+    """Render loop for Gradio"""
+    setup_state(state)
+    return "Setup completed"
+
+def validate_auth(provider: APIProvider, api_key: str | None):
+    if provider == APIProvider.ANTHROPIC:
+        if not api_key:
+            return "Enter your Anthropic API key to continue."
+    if provider == APIProvider.BEDROCK:
+        import boto3
+
+        if not boto3.Session().get_credentials():
+            return "You must have AWS credentials set up to use the Bedrock API."
+    if provider == APIProvider.VERTEX:
+        import google.auth
+        from google.auth.exceptions import DefaultCredentialsError
+
+        if not os.environ.get("CLOUD_ML_REGION"):
+            return "Set the CLOUD_ML_REGION environment variable to use the Vertex API."
+        try:
+            google.auth.default(scopes=["https://www.googleapis.com/auth/cloud-platform"])
+        except DefaultCredentialsError:
+            return "Your google cloud credentials are not set up correctly."
+
+def load_from_storage(filename: str) -> str | None:
+    """Load data from a file in the storage directory."""
+    try:
+        file_path = CONFIG_DIR / filename
+        if file_path.exists():
+            data = file_path.read_text().strip()
+            if data:
+                return data
+    except Exception as e:
+        print(f"Debug: Error loading {filename}: {e}")
+    return None
+
+def save_to_storage(filename: str, data: str) -> None:
+    """Save data to a file in the storage directory."""
+    try:
+        CONFIG_DIR.mkdir(parents=True, exist_ok=True)
+        file_path = CONFIG_DIR / filename
+        file_path.write_text(data)
+        # Ensure only user can read/write the file
+        file_path.chmod(0o600)
+    except Exception as e:
+        print(f"Debug: Error saving {filename}: {e}")
+
+def _api_response_callback(response: APIResponse[BetaMessage], response_state: dict):
+    response_id = datetime.now().isoformat()
+    response_state[response_id] = response
+
+def _tool_output_callback(tool_output: ToolResult, tool_id: str, tool_state: dict):
+    tool_state[tool_id] = tool_output
+
+def chatbot_output_callback(message, chatbot_state, hide_images=False, sender="bot"):
+    def _render_message(message: str | BetaTextBlock | BetaToolUseBlock | ToolResult, hide_images=False):
+    
+        print(f"_render_message: {str(message)[:100]}")
+        
+        if isinstance(message, str):
+            return message
+        
+        is_tool_result = not isinstance(message, str) and (
+            isinstance(message, ToolResult)
+            or message.__class__.__name__ == "ToolResult"
+        )
+        if not message or (
+            is_tool_result
+            and hide_images
+            and not hasattr(message, "error")
+            and not hasattr(message, "output")
+        ):  # return None if hide_images is True
+            return
+        # render tool result
+        if is_tool_result:
+            message = cast(ToolResult, message)
+            if message.output:
+                return message.output
+            if message.error:
+                return f"Error: {message.error}"
+            if message.base64_image and not hide_images:
+                # somehow can't display via gr.Image
+                # image_data = base64.b64decode(message.base64_image)
+                # return gr.Image(value=Image.open(io.BytesIO(image_data)))
+                return f'<img src="data:image/png;base64,{message.base64_image}">'
+
+        elif isinstance(message, BetaTextBlock) or isinstance(message, TextBlock):
+            return f"Analysis: {message.text}"
+        elif isinstance(message, BetaToolUseBlock) or isinstance(message, ToolUseBlock):
+            # return f"Tool Use: {message.name}\nInput: {message.input}"
+            return f"Next I will perform the following action: {message.input}"
+        else:  
+            return message
+
+    def _truncate_string(s, max_length=500):
+        """Truncate long strings for concise printing."""
+        if isinstance(s, str) and len(s) > max_length:
+            return s[:max_length] + "..."
+        return s
+    # processing Anthropic messages
+    message = _render_message(message, hide_images)
+    
+    if sender == "bot":
+        chatbot_state.append((None, message))
+    else:
+        chatbot_state.append((message, None))
+    
+    # Create a concise version of the chatbot state for printing
+    concise_state = [(_truncate_string(user_msg), _truncate_string(bot_msg))
+                        for user_msg, bot_msg in chatbot_state]
+    # print(f"chatbot_output_callback chatbot_state: {concise_state} (truncated)")
+
+def valid_params(user_input, state):
+    """Validate all requirements and return a list of error messages."""
+    errors = []
+    
+    for server_name, url in [('Windows Host', 'localhost:5000'), ('OmniParser Server', args.omniparser_server_url)]:
+        try:
+            url = f'http://{url}/probe'
+            response = requests.get(url, timeout=3)
+            if response.status_code != 200:
+                errors.append(f"{server_name} is not responding")
+        except RequestException as e:
+            errors.append(f"{server_name} is not responding")
+    
+    if not state["api_key"].strip():
+        errors.append("LLM API Key is not set")
+
+    if not user_input:
+        errors.append("no computer use request provided")
+    
+    return errors
+
+def process_input(user_input, state):
+    # Reset the stop flag
+    if state["stop"]:
+        state["stop"] = False
+
+    errors = valid_params(user_input, state)
+    if errors:
+        raise gr.Error("Validation errors: " + ", ".join(errors))
+    
+    # Append the user message to state["messages"]
+    state["messages"].append(
+        {
+            "role": Sender.USER,
+            "content": [TextBlock(type="text", text=user_input)],
+        }
+    )
+
+    # Append the user's message to chatbot_messages with None for the assistant's reply
+    state['chatbot_messages'].append((user_input, None))
+    yield state['chatbot_messages']  # Yield to update the chatbot UI with the user's message
+
+    print("state")
+    print(state)
+
+    # Run sampling_loop_sync with the chatbot_output_callback
+    for loop_msg in sampling_loop_sync(
+        model=state["model"],
+        provider=state["provider"],
+        messages=state["messages"],
+        output_callback=partial(chatbot_output_callback, chatbot_state=state['chatbot_messages'], hide_images=False),
+        tool_output_callback=partial(_tool_output_callback, tool_state=state["tools"]),
+        api_response_callback=partial(_api_response_callback, response_state=state["responses"]),
+        api_key=state["api_key"],
+        only_n_most_recent_images=state["only_n_most_recent_images"],
+        max_tokens=16384,
+        omniparser_url=args.omniparser_server_url
+    ):  
+        if loop_msg is None or state.get("stop"):
+            yield state['chatbot_messages']
+            print("End of task. Close the loop.")
+            break
+            
+        yield state['chatbot_messages']  # Yield the updated chatbot_messages to update the chatbot UI
+
+def stop_app(state):
+    state["stop"] = True
+    return "App stopped"
+
+def get_header_image_base64():
+    try:
+        # Get the absolute path to the image relative to this script
+        script_dir = Path(__file__).parent
+        image_path = script_dir.parent.parent / "imgs" / "header_bar_thin.png"
+        
+        with open(image_path, "rb") as image_file:
+            encoded_string = base64.b64encode(image_file.read()).decode()
+            return f'data:image/png;base64,{encoded_string}'
+    except Exception as e:
+        print(f"Failed to load header image: {e}")
+        return None
+
+with gr.Blocks(theme=gr.themes.Default()) as demo:
+    gr.HTML("""
+        <style>
+        .no-padding {
+            padding: 0 !important;
+        }
+        .no-padding > div {
+            padding: 0 !important;
+        }
+        .markdown-text p {
+            font-size: 18px;  /* Adjust the font size as needed */
+        }
+        </style>
+    """)
+    state = gr.State({})
+    
+    setup_state(state.value)
+    
+    header_image = get_header_image_base64()
+    if header_image:
+        gr.HTML(f'<img src="{header_image}" alt="OmniTool Header" width="100%">', elem_classes="no-padding")
+        gr.HTML('<h1 style="text-align: center; font-weight: normal;">Omni<span style="font-weight: bold;">Tool</span></h1>')
+    else:
+        gr.Markdown("# OmniTool")
+
+    if not os.getenv("HIDE_WARNING", False):
+        gr.Markdown(INTRO_TEXT, elem_classes="markdown-text")
+
+
+    with gr.Accordion("Settings", open=True): 
+        with gr.Row():
+            with gr.Column():
+                model = gr.Dropdown(
+                    label="Model",
+                    choices=["omniparser + gpt-4o", "omniparser + o1", "omniparser + o3-mini", "omniparser + R1", "omniparser + qwen2.5vl", "claude-3-5-sonnet-20241022", "omniparser + gpt-4o-orchestrated", "omniparser + o1-orchestrated", "omniparser + o3-mini-orchestrated", "omniparser + R1-orchestrated", "omniparser + qwen2.5vl-orchestrated"],
+                    value="omniparser + gpt-4o",
+                    interactive=True,
+                )
+            with gr.Column():
+                only_n_images = gr.Slider(
+                    label="N most recent screenshots",
+                    minimum=0,
+                    maximum=10,
+                    step=1,
+                    value=2,
+                    interactive=True
+                )
+        with gr.Row():
+            with gr.Column(1):
+                provider = gr.Dropdown(
+                    label="API Provider",
+                    choices=[option.value for option in APIProvider],
+                    value="openai",
+                    interactive=False,
+                )
+            with gr.Column(2):
+                api_key = gr.Textbox(
+                    label="API Key",
+                    type="password",
+                    value=state.value.get("api_key", ""),
+                    placeholder="Paste your API key here",
+                    interactive=True,
+                )
+
+    with gr.Row():
+        with gr.Column(scale=8):
+            chat_input = gr.Textbox(show_label=False, placeholder="Type a message to send to Omniparser + X ...", container=False)
+        with gr.Column(scale=1, min_width=50):
+            submit_button = gr.Button(value="Send", variant="primary")
+        with gr.Column(scale=1, min_width=50):
+            stop_button = gr.Button(value="Stop", variant="secondary")
+
+    with gr.Row():
+        with gr.Column(scale=2):
+            chatbot = gr.Chatbot(label="Chatbot History", autoscroll=True, height=580)
+        with gr.Column(scale=3):
+            iframe = gr.HTML(
+                f'<iframe src="http://{args.windows_host_url}/vnc.html?view_only=1&autoconnect=1&resize=scale" width="100%" height="580" allow="fullscreen"></iframe>',
+                container=False,
+                elem_classes="no-padding"
+            )
+
+    def update_model(model_selection, state):
+        state["model"] = model_selection
+        print(f"Model updated to: {state['model']}")
+        
+        if model_selection == "claude-3-5-sonnet-20241022":
+            provider_choices = [option.value for option in APIProvider if option.value != "openai"]
+        elif model_selection in set(["omniparser + gpt-4o", "omniparser + o1", "omniparser + o3-mini", "omniparser + gpt-4o-orchestrated", "omniparser + o1-orchestrated", "omniparser + o3-mini-orchestrated"]):
+            provider_choices = ["openai"]
+        elif model_selection == "omniparser + R1":
+            provider_choices = ["groq"]
+        elif model_selection == "omniparser + qwen2.5vl":
+            provider_choices = ["dashscope"]
+        else:
+            provider_choices = [option.value for option in APIProvider]
+        default_provider_value = provider_choices[0]
+
+        provider_interactive = len(provider_choices) > 1
+        api_key_placeholder = f"{default_provider_value.title()} API Key"
+
+        # Update state
+        state["provider"] = default_provider_value
+        state["api_key"] = state.get(f"{default_provider_value}_api_key", "")
+
+        # Calls to update other components UI
+        provider_update = gr.update(
+            choices=provider_choices,
+            value=default_provider_value,
+            interactive=provider_interactive
+        )
+        api_key_update = gr.update(
+            placeholder=api_key_placeholder,
+            value=state["api_key"]
+        )
+
+        return provider_update, api_key_update
+
+    def update_only_n_images(only_n_images_value, state):
+        state["only_n_most_recent_images"] = only_n_images_value
+   
+    def update_provider(provider_value, state):
+        # Update state
+        state["provider"] = provider_value
+        state["api_key"] = state.get(f"{provider_value}_api_key", "")
+        
+        # Calls to update other components UI
+        api_key_update = gr.update(
+            placeholder=f"{provider_value.title()} API Key",
+            value=state["api_key"]
+        )
+        return api_key_update
+                
+    def update_api_key(api_key_value, state):
+        state["api_key"] = api_key_value
+        state[f'{state["provider"]}_api_key'] = api_key_value
+
+    def clear_chat(state):
+        # Reset message-related state
+        state["messages"] = []
+        state["responses"] = {}
+        state["tools"] = {}
+        state['chatbot_messages'] = []
+        return state['chatbot_messages']
+
+    model.change(fn=update_model, inputs=[model, state], outputs=[provider, api_key])
+    only_n_images.change(fn=update_only_n_images, inputs=[only_n_images, state], outputs=None)
+    provider.change(fn=update_provider, inputs=[provider, state], outputs=api_key)
+    api_key.change(fn=update_api_key, inputs=[api_key, state], outputs=None)
+    chatbot.clear(fn=clear_chat, inputs=[state], outputs=[chatbot])
+
+    submit_button.click(process_input, [chat_input, state], chatbot)
+    stop_button.click(stop_app, [state], None)
+    
+if __name__ == "__main__":
+    demo.launch(server_name="0.0.0.0", server_port=7888)

OmniParser/omnitool/gradio/app_new.py

@@ -0,0 +1,760 @@
+"""
+The app contains:
+- a new UI for the OmniParser AI Agent.
+- 
+python app_new.py --windows_host_url localhost:8006 --omniparser_server_url localhost:8000
+"""
+
+import os
+import io
+import shutil
+import mimetypes
+from datetime import datetime
+from enum import StrEnum
+from functools import partial
+from pathlib import Path
+from typing import cast, List, Optional
+import argparse
+import gradio as gr
+from anthropic import APIResponse
+from anthropic.types import TextBlock
+from anthropic.types.beta import BetaMessage, BetaTextBlock, BetaToolUseBlock
+from anthropic.types.tool_use_block import ToolUseBlock
+from loop import (
+    APIProvider,
+    sampling_loop_sync,
+)
+from tools import ToolResult
+import requests
+from requests.exceptions import RequestException
+import base64
+
+CONFIG_DIR = Path("~/.anthropic").expanduser()
+API_KEY_FILE = CONFIG_DIR / "api_key"
+
+INTRO_TEXT = '''
+<div style="text-align: center; margin-bottom: 10px;">
+    <h2>OmniParser AI Agent</h2>
+    <p>Turn any vision-language model into an AI agent. We currently support <b>OpenAI (4o/o1/o3-mini), DeepSeek (R1), Qwen (2.5VL) or Anthropic Computer Use (Sonnet)</b>.</p>
+    <p>Type a message and press send to start OmniTool. Press stop to pause, and press the trash icon in the chat to clear the message history.</p>
+    <p>You can also upload files for analysis using the file upload section.</p>
+</div>
+'''
+
+def parse_arguments():
+    parser = argparse.ArgumentParser(description="Gradio App")
+    parser.add_argument("--windows_host_url", type=str, default='localhost:8006')
+    parser.add_argument("--omniparser_server_url", type=str, default="localhost:8000")
+    parser.add_argument("--run_folder", type=str, default="./tmp/outputs")
+    return parser.parse_args()
+args = parse_arguments()
+
+# Update upload folder from args if provided
+RUN_FOLDER = Path(os.path.join(args.run_folder, datetime.now().strftime('%Y%m%d_%H%M')))
+RUN_FOLDER.mkdir(parents=True, exist_ok=True)
+
+class Sender(StrEnum):
+    USER = "user"
+    BOT = "assistant"
+    TOOL = "tool"
+
+
+def load_existing_files():
+    """Load all existing files from the uploads folder"""
+    files = []
+    if RUN_FOLDER.exists():
+        for file_path in RUN_FOLDER.iterdir():
+            if file_path.is_file():
+                files.append(str(file_path))
+    return files
+
+def setup_state(state):
+    if "messages" not in state:
+        state["messages"] = []
+    if "model" not in state:
+        state["model"] = "omniparser + gpt-4o-orchestrated"
+    if "provider" not in state:
+        state["provider"] = "openai"
+    if "openai_api_key" not in state:  # Fetch API keys from environment variables
+        state["openai_api_key"] = os.getenv("OPENAI_API_KEY", "")
+    if "anthropic_api_key" not in state:
+        state["anthropic_api_key"] = os.getenv("ANTHROPIC_API_KEY", "")
+    if "api_key" not in state:
+        state["api_key"] = ""
+    if "auth_validated" not in state:
+        state["auth_validated"] = False
+    if "responses" not in state:
+        state["responses"] = {}
+    if "tools" not in state:
+        state["tools"] = {}
+    if "only_n_most_recent_images" not in state:
+        state["only_n_most_recent_images"] = 2
+    if 'chatbot_messages' not in state:
+        state['chatbot_messages'] = []
+    if 'stop' not in state:
+        state['stop'] = False
+    if 'uploaded_files' not in state:
+        state['uploaded_files'] = []  # Start with an empty list instead of loading existing files
+
+async def main(state):
+    """Render loop for Gradio"""
+    setup_state(state)
+    return "Setup completed"
+
+def validate_auth(provider: APIProvider, api_key: str | None):
+    if provider == APIProvider.ANTHROPIC:
+        if not api_key:
+            return "Enter your Anthropic API key to continue."
+    if provider == APIProvider.BEDROCK:
+        import boto3
+
+        if not boto3.Session().get_credentials():
+            return "You must have AWS credentials set up to use the Bedrock API."
+    if provider == APIProvider.VERTEX:
+        import google.auth
+        from google.auth.exceptions import DefaultCredentialsError
+
+        if not os.environ.get("CLOUD_ML_REGION"):
+            return "Set the CLOUD_ML_REGION environment variable to use the Vertex API."
+        try:
+            google.auth.default(scopes=["https://www.googleapis.com/auth/cloud-platform"])
+        except DefaultCredentialsError:
+            return "Your google cloud credentials are not set up correctly."
+
+def load_from_storage(filename: str) -> str | None:
+    """Load data from a file in the storage directory."""
+    try:
+        file_path = CONFIG_DIR / filename
+        if file_path.exists():
+            data = file_path.read_text().strip()
+            if data:
+                return data
+    except Exception as e:
+        print(f"Debug: Error loading {filename}: {e}")
+    return None
+
+def save_to_storage(filename: str, data: str) -> None:
+    """Save data to a file in the storage directory."""
+    try:
+        CONFIG_DIR.mkdir(parents=True, exist_ok=True)
+        file_path = CONFIG_DIR / filename
+        file_path.write_text(data)
+        # Ensure only user can read/write the file
+        file_path.chmod(0o600)
+    except Exception as e:
+        print(f"Debug: Error saving {filename}: {e}")
+
+def _api_response_callback(response: APIResponse[BetaMessage], response_state: dict):
+    response_id = datetime.now().isoformat()
+    response_state[response_id] = response
+
+def _tool_output_callback(tool_output: ToolResult, tool_id: str, tool_state: dict):
+    tool_state[tool_id] = tool_output
+
+def chatbot_output_callback(message, chatbot_state, hide_images=False, sender="bot"):
+    def _render_message(message: str | BetaTextBlock | BetaToolUseBlock | ToolResult, hide_images=False):
+    
+        print(f"_render_message: {str(message)[:100]}")
+        
+        if isinstance(message, str):
+            return message
+        
+        is_tool_result = not isinstance(message, str) and (
+            isinstance(message, ToolResult)
+            or message.__class__.__name__ == "ToolResult"
+        )
+        if not message or (
+            is_tool_result
+            and hide_images
+            and not hasattr(message, "error")
+            and not hasattr(message, "output")
+        ):  # return None if hide_images is True
+            return
+        # render tool result
+        if is_tool_result:
+            message = cast(ToolResult, message)
+            if message.output:
+                return message.output
+            if message.error:
+                return f"Error: {message.error}"
+            if message.base64_image and not hide_images:
+                # somehow can't display via gr.Image
+                # image_data = base64.b64decode(message.base64_image)
+                # return gr.Image(value=Image.open(io.BytesIO(image_data)))
+                return f'<img src="data:image/png;base64,{message.base64_image}">'
+
+        elif isinstance(message, BetaTextBlock) or isinstance(message, TextBlock):
+            # Format reasoning text in a collapsible dropdown
+            return f"Next step Reasoning: {message.text}"
+            # reasoning_text = message.text
+            # return f'''
+            # <details>
+            #     <summary><Current Step Reasoning (click to expand):</summary>
+            #     <div style="padding: 10px; background-color: #f8f9fa; border-radius: 5px; margin-top: 5px;">
+            #         <pre>{reasoning_text}</pre>
+            #     </div>
+            # </details>
+            # '''
+        elif isinstance(message, BetaToolUseBlock) or isinstance(message, ToolUseBlock):
+            # return f"Next I will perform the following action: {message.input}"
+            return None
+        else:  
+            return message
+
+    def _truncate_string(s, max_length=500):
+        """Truncate long strings for concise printing."""
+        if isinstance(s, str) and len(s) > max_length:
+            return s[:max_length] + "..."
+        return s
+    # processing Anthropic messages
+    message = _render_message(message, hide_images)
+    
+    if sender == "bot":
+        chatbot_state.append((None, message))
+    else:
+        chatbot_state.append((message, None))
+    
+    # Create a concise version of the chatbot state for printing
+    concise_state = [(_truncate_string(user_msg), _truncate_string(bot_msg))
+                        for user_msg, bot_msg in chatbot_state]
+    # print(f"chatbot_output_callback chatbot_state: {concise_state} (truncated)")
+
+def valid_params(user_input, state):
+    """Validate all requirements and return a list of error messages."""
+    errors = []
+    
+    for server_name, url in [('Windows Host', 'localhost:5000'), ('OmniParser Server', args.omniparser_server_url)]:
+        try:
+            url = f'http://{url}/probe'
+            response = requests.get(url, timeout=3)
+            if response.status_code != 200:
+                errors.append(f"{server_name} is not responding")
+        except RequestException as e:
+            errors.append(f"{server_name} is not responding")
+    
+    if not state["api_key"].strip():
+        errors.append("LLM API Key is not set")
+
+    if not user_input:
+        errors.append("no computer use request provided")
+    
+    return errors
+
+def process_input(user_input, state):
+    # Reset the stop flag
+    if state["stop"]:
+        state["stop"] = False
+
+    errors = valid_params(user_input, state)
+    if errors:
+        raise gr.Error("Validation errors: " + ", ".join(errors))
+    
+    # Append the user message to state["messages"]
+    state["messages"].append(
+        {
+            "role": Sender.USER,
+            "content": [TextBlock(type="text", text=user_input)],
+        }
+    )
+
+    # Append the user's message to chatbot_messages with None for the assistant's reply
+    state['chatbot_messages'].append((user_input, None))
+    yield state['chatbot_messages'], gr.update()  # Yield to update the chatbot UI with the user's message
+
+    print("state")
+    print(state)
+
+    # Run sampling_loop_sync with the chatbot_output_callback
+    for loop_msg in sampling_loop_sync(
+        model=state["model"],
+        provider=state["provider"],
+        messages=state["messages"],
+        output_callback=partial(chatbot_output_callback, chatbot_state=state['chatbot_messages'], hide_images=False),
+        tool_output_callback=partial(_tool_output_callback, tool_state=state["tools"]),
+        api_response_callback=partial(_api_response_callback, response_state=state["responses"]),
+        api_key=state["api_key"],
+        only_n_most_recent_images=state["only_n_most_recent_images"],
+        max_tokens=16384,
+        omniparser_url=args.omniparser_server_url,
+        save_folder=str(RUN_FOLDER)
+    ):  
+        if loop_msg is None or state.get("stop"):
+            # Detect and add new files to the state
+            file_choices_update = detect_new_files(state)
+            yield state['chatbot_messages'], file_choices_update
+            print("End of task. Close the loop.")
+            break
+            
+        yield state['chatbot_messages'], gr.update()  # Yield the updated chatbot_messages to update the chatbot UI
+    
+    # Final detection of new files
+    file_choices_update = detect_new_files(state)
+    yield state['chatbot_messages'], file_choices_update
+
+def stop_app(state):
+    state["stop"] = True
+    return "App stopped"
+
+def get_header_image_base64():
+    try:
+        # Get the absolute path to the image relative to this script
+        script_dir = Path(__file__).parent
+        image_path = script_dir.parent.parent / "imgs" / "header_bar_thin.png"
+        
+        with open(image_path, "rb") as image_file:
+            encoded_string = base64.b64encode(image_file.read()).decode()
+            return f'data:image/png;base64,{encoded_string}'
+    except Exception as e:
+        print(f"Failed to load header image: {e}")
+        return None
+
+def get_file_viewer_html(file_path=None):
+    """Generate HTML to view a file based on its type"""
+    if not file_path:
+        # Return the VNC viewer iframe
+        return f'<iframe src="http://{args.windows_host_url}/vnc.html?view_only=1&autoconnect=1&resize=scale" width="100%" height="580" allow="fullscreen"></iframe>'
+    
+    file_path = Path(file_path)
+    if not file_path.exists():
+        return f'<div class="error-message">File not found: {file_path.name}</div>'
+    
+    # Determine the file type
+    mime_type, _ = mimetypes.guess_type(file_path)
+    file_type = mime_type.split('/')[0] if mime_type else 'unknown'
+    file_extension = file_path.suffix.lower()
+    
+    # Handle different file types
+    if file_type == 'image':
+        # For images, display them directly
+        with open(file_path, "rb") as image_file:
+            encoded_string = base64.b64encode(image_file.read()).decode()
+            return f'<div class="file-viewer"><h3>{file_path.name}</h3><img src="data:{mime_type};base64,{encoded_string}" style="max-width:100%; max-height:500px;"></div>'
+    
+    elif file_extension in ['.txt', '.py', '.js', '.html', '.css', '.json', '.md', '.csv'] or file_type == 'text':
+        # For text files, display the content with syntax highlighting for code
+        try:
+            content = file_path.read_text(errors='replace')  # Use 'replace' to handle encoding issues
+            # Escape HTML characters
+            content = content.replace('&', '&amp;').replace('<', '&lt;').replace('>', '&gt;')
+            
+            # Add syntax highlighting class based on file extension
+            highlight_class = ""
+            if file_extension == '.py':
+                highlight_class = "language-python"
+            elif file_extension == '.js':
+                highlight_class = "language-javascript"
+            elif file_extension == '.html':
+                highlight_class = "language-html"
+            elif file_extension == '.css':
+                highlight_class = "language-css"
+            elif file_extension == '.json':
+                highlight_class = "language-json"
+            
+            return f'''
+            <div class="file-viewer">
+                <h3>{file_path.name}</h3>
+                <pre class="{highlight_class}" style="background-color: #f5f5f5; padding: 10px; border-radius: 5px; overflow: auto; max-height: 500px; white-space: pre-wrap;"><code>{content}</code></pre>
+                <script>
+                    // Add basic syntax highlighting with CSS
+                    if (document.querySelector('.language-python')) {{
+                        const keywords = ['def', 'class', 'import', 'from', 'return', 'if', 'else', 'elif', 'for', 'while', 'try', 'except', 'with', 'as', 'in', 'not', 'and', 'or', 'True', 'False', 'None'];
+                        const code = document.querySelector('.language-python code');
+                        let html = code.innerHTML;
+                        keywords.forEach(keyword => {{
+                            const regex = new RegExp('\\\\b' + keyword + '\\\\b', 'g');
+                            html = html.replace(regex, `<span style="color: #0000FF; font-weight: bold;">$&</span>`);
+                        }});
+                        // Highlight strings
+                        html = html.replace(/(["'])(?:(?=(\\\\?))\2.)*?\1/g, '<span style="color: #008000;">$&</span>');
+                        // Highlight comments
+                        html = html.replace(/(#.*)$/gm, '<span style="color: #808080;">$1</span>');
+                        code.innerHTML = html;
+                    }}
+                </script>
+            </div>
+            '''
+        except UnicodeDecodeError:
+            return f'<div class="error-message">Cannot display binary file: {file_path.name}</div>'
+    
+    elif file_type == 'video':
+        # For videos, use video tag
+        with open(file_path, "rb") as video_file:
+            encoded_string = base64.b64encode(video_file.read()).decode()
+            return f'''
+            <div class="file-viewer">
+                <h3>{file_path.name}</h3>
+                <video controls style="max-width:100%; max-height:500px;">
+                    <source src="data:{mime_type};base64,{encoded_string}" type="{mime_type}">
+                    Your browser does not support the video tag.
+                </video>
+            </div>
+            '''
+    
+    elif file_type == 'audio':
+        # For audio, use audio tag
+        with open(file_path, "rb") as audio_file:
+            encoded_string = base64.b64encode(audio_file.read()).decode()
+            return f'''
+            <div class="file-viewer">
+                <h3>{file_path.name}</h3>
+                <audio controls>
+                    <source src="data:{mime_type};base64,{encoded_string}" type="{mime_type}">
+                    Your browser does not support the audio tag.
+                </audio>
+            </div>
+            '''
+    
+    elif file_extension == '.pdf':
+        # For PDFs, embed them using an iframe with base64 data
+        try:
+            with open(file_path, "rb") as pdf_file:
+                encoded_string = base64.b64encode(pdf_file.read()).decode()
+                return f'''
+                <div class="file-viewer">
+                    <h3>{file_path.name}</h3>
+                    <iframe src="data:application/pdf;base64,{encoded_string}" width="100%" height="500px" style="border: none;"></iframe>
+                </div>
+                '''
+        except Exception as e:
+            return f'<div class="error-message">Error displaying PDF: {str(e)}</div>'
+    
+    else:
+        # For other file types, show info but can't display
+        size_kb = file_path.stat().st_size / 1024
+        return f'<div class="file-viewer"><h3>{file_path.name}</h3><p>File type: {mime_type or "Unknown"}</p><p>Size: {size_kb:.2f} KB</p><p>This file type cannot be displayed in the browser.</p></div>'
+
+def handle_file_upload(files, state):
+    """Handle file uploads and store them in the upload directory"""
+    if not files:
+        return gr.update(choices=[])
+    
+    file_choices = []
+    
+    for file in files:
+        # Get the file name and create a path in the upload directory
+        file_name = Path(file.name).name
+        file_path = RUN_FOLDER / file_name
+        
+        # Save the file
+        shutil.copy(file.name, file_path)
+        
+        # Add to the list of uploaded files
+        file_path_str = str(file_path)
+        file_choices.append((file_name, file_path_str))
+        
+        # Add to state
+        if file_path_str not in state['uploaded_files']:
+            state['uploaded_files'].append(file_path_str)
+    
+    # Update the view file dropdown with all uploaded files
+    all_file_choices = [(Path(path).name, path) for path in state['uploaded_files']]
+    
+    return gr.update(choices=all_file_choices)
+
+def toggle_view(view_mode, file_path=None, state=None):
+    """Toggle between OmniTool Computer view and file viewer"""
+    # If switching to File Viewer mode, detect and add new files to the state
+    file_choices_update = gr.update()
+    if view_mode == "File Viewer" and state is not None:
+        file_choices_update = detect_new_files(state)
+    
+    # Return the appropriate view
+    if view_mode == "OmniTool Computer":
+        return get_file_viewer_html(), file_choices_update  # This returns the VNC iframe
+    else:  # File Viewer mode
+        if file_path:
+            return get_file_viewer_html(file_path), file_choices_update
+        else:
+            return get_file_viewer_html(), file_choices_update  # Default to VNC if no file selected
+
+def detect_new_files(state):
+    """Detect new files in the uploads folder and add them to the state"""
+    new_files_count = 0
+    if RUN_FOLDER.exists():
+        current_files = set(state['uploaded_files'])
+        for file_path in RUN_FOLDER.iterdir():
+            if file_path.is_file():
+                file_path_str = str(file_path)
+                if file_path_str not in current_files:
+                    # This is a new file not yet in the state
+                    state['uploaded_files'].append(file_path_str)
+                    new_files_count += 1
+                    print(f"Added new file to state: {file_path_str}")
+    
+    # Return updated file choices
+    file_choices = [(Path(path).name, path) for path in state['uploaded_files']]
+    print(f"Detected {new_files_count} new files. Total files in state: {len(state['uploaded_files'])}")
+    return gr.update(choices=file_choices)
+
+def refresh_files(state):
+    """Refresh the list of files from the current session and detect new files"""
+    return detect_new_files(state)
+
+def auto_refresh_files(state):
+    """Automatically refresh the list of files from the current session and detect new files"""
+    return detect_new_files(state)
+
+with gr.Blocks(theme=gr.themes.Default()) as demo:
+    gr.HTML("""
+        <style>
+        .no-padding {
+            padding: 0 !important;
+        }
+        .no-padding > div {
+            padding: 0 !important;
+        }
+        .markdown-text p {
+            font-size: 18px;  /* Adjust the font size as needed */
+        }
+        </style>
+    """)
+    state = gr.State({})
+    
+    setup_state(state.value)
+    
+    header_image = get_header_image_base64()
+    if header_image:
+        gr.HTML(f'<img src="{header_image}" alt="OmniTool Header" width="100%">', elem_classes="no-padding")
+        gr.HTML('<h1 style="text-align: center; font-weight: normal; margin-bottom: 20px;">Omni<span style="font-weight: bold;">Tool</span></h1>')
+    else:
+        gr.Markdown("# OmniTool", elem_classes="text-center")
+
+    if not os.getenv("HIDE_WARNING", False):
+        gr.HTML(INTRO_TEXT, elem_classes="markdown-text")
+
+    with gr.Accordion("Settings", open=True, elem_classes="accordion-header"): 
+        with gr.Row():
+            with gr.Column():
+                model = gr.Dropdown(
+                    label="Model",
+                    choices=["omniparser + gpt-4o", "omniparser + o1", "omniparser + o3-mini", "omniparser + R1", "omniparser + qwen2.5vl", "claude-3-5-sonnet-20241022", "omniparser + gpt-4o-orchestrated", "omniparser + o1-orchestrated", "omniparser + o3-mini-orchestrated", "omniparser + R1-orchestrated", "omniparser + qwen2.5vl-orchestrated"],
+                    value="omniparser + gpt-4o-orchestrated",
+                    interactive=True,
+                    container=True
+                )
+            with gr.Column():
+                only_n_images = gr.Slider(
+                    label="N most recent screenshots",
+                    minimum=0,
+                    maximum=10,
+                    step=1,
+                    value=2,
+                    interactive=True
+                )
+        with gr.Row():
+            with gr.Column(1):
+                provider = gr.Dropdown(
+                    label="API Provider",
+                    choices=[option.value for option in APIProvider],
+                    value="openai",
+                    interactive=False,
+                    container=True
+                )
+            with gr.Column(2):
+                api_key = gr.Textbox(
+                    label="API Key",
+                    type="password",
+                    value=state.value.get("api_key", ""),
+                    placeholder="Paste your API key here",
+                    interactive=True,
+                    container=True
+                )
+
+    # File Upload Section
+    with gr.Accordion("File Upload & Management", open=True, elem_classes="accordion-header"):
+        with gr.Row():
+            with gr.Column():
+                file_upload = gr.File(
+                    label="Upload Files",
+                    file_count="multiple",
+                    type="filepath",
+                    elem_classes="file-upload-area"
+                )
+            with gr.Column():
+                with gr.Row():
+                    upload_button = gr.Button("Upload Files", variant="primary", elem_classes="primary-button")
+                    refresh_button = gr.Button("Refresh Files", variant="secondary", elem_classes="secondary-button")
+        
+        with gr.Row():
+            # Initialize file choices as an empty list
+            view_file_dropdown = gr.Dropdown(
+                label="View File",
+                choices=[],
+                interactive=True,
+                container=True
+            )
+            view_toggle = gr.Radio(
+                label="Display Mode",
+                choices=["OmniTool Computer", "File Viewer"],
+                value="OmniTool Computer",
+                interactive=True
+            )
+
+    with gr.Row():
+        with gr.Column(scale=8):
+            chat_input = gr.Textbox(
+                show_label=False, 
+                placeholder="Type a message to send to Omniparser + X ...", 
+                container=False
+            )
+        with gr.Column(scale=1, min_width=50):
+            submit_button = gr.Button(value="Send", variant="primary", elem_classes="primary-button")
+        with gr.Column(scale=1, min_width=50):
+            stop_button = gr.Button(value="Stop", variant="secondary", elem_classes="secondary-button")
+
+    with gr.Row():
+        with gr.Column(scale=2):
+            chatbot = gr.Chatbot(
+                label="Chatbot History", 
+                autoscroll=True, 
+                height=580,
+                avatar_images=("👤", "🤖")
+            )
+        with gr.Column(scale=3):
+            display_area = gr.HTML(
+                get_file_viewer_html(),
+                elem_classes="no-padding"
+            )
+
+    def update_model(model_selection, state):
+        state["model"] = model_selection
+        print(f"Model updated to: {state['model']}")
+        
+        if model_selection == "claude-3-5-sonnet-20241022":
+            provider_choices = [option.value for option in APIProvider if option.value != "openai"]
+        elif model_selection in set(["omniparser + gpt-4o", "omniparser + o1", "omniparser + o3-mini", "omniparser + gpt-4o-orchestrated", "omniparser + o1-orchestrated", "omniparser + o3-mini-orchestrated"]):
+            provider_choices = ["openai"]
+        elif model_selection == "omniparser + R1":
+            provider_choices = ["groq"]
+        elif model_selection == "omniparser + qwen2.5vl":
+            provider_choices = ["dashscope"]
+        else:
+            provider_choices = [option.value for option in APIProvider]
+        default_provider_value = provider_choices[0]
+
+        provider_interactive = len(provider_choices) > 1
+        api_key_placeholder = f"{default_provider_value.title()} API Key"
+
+        # Update state
+        state["provider"] = default_provider_value
+        state["api_key"] = state.get(f"{default_provider_value}_api_key", "")
+
+        # Calls to update other components UI
+        provider_update = gr.update(
+            choices=provider_choices,
+            value=default_provider_value,
+            interactive=provider_interactive
+        )
+        api_key_update = gr.update(
+            placeholder=api_key_placeholder,
+            value=state["api_key"]
+        )
+
+        return provider_update, api_key_update
+
+    def update_only_n_images(only_n_images_value, state):
+        state["only_n_most_recent_images"] = only_n_images_value
+   
+    def update_provider(provider_value, state):
+        # Update state
+        state["provider"] = provider_value
+        state["api_key"] = state.get(f"{provider_value}_api_key", "")
+        
+        # Calls to update other components UI
+        api_key_update = gr.update(
+            placeholder=f"{provider_value.title()} API Key",
+            value=state["api_key"]
+        )
+        return api_key_update
+                
+    def update_api_key(api_key_value, state):
+        state["api_key"] = api_key_value
+        state[f'{state["provider"]}_api_key'] = api_key_value
+
+    def clear_chat(state):
+        # Reset message-related state
+        state["messages"] = []
+        state["responses"] = {}
+        state["tools"] = {}
+        state['chatbot_messages'] = []
+        return state['chatbot_messages']
+
+    def view_file(file_path, view_mode):
+        """Generate HTML to view the selected file if in File Viewer mode"""
+        if view_mode == "File Viewer" and file_path:
+            return get_file_viewer_html(file_path)
+        elif view_mode == "OmniTool Computer":
+            return get_file_viewer_html()  # Return VNC viewer
+        else:
+            return display_area.value  # Keep current display
+
+    def update_view_file_dropdown(uploaded_files):
+        """Update the view file dropdown when uploaded files change"""
+        if not uploaded_files:
+            return gr.update(choices=[])
+        
+        file_choices = [(Path(path).name, path) for path in uploaded_files]
+        return gr.update(choices=file_choices)
+
+    def reset_view():
+        """Reset the view to the VNC viewer"""
+        return get_file_viewer_html()
+
+    model.change(fn=update_model, inputs=[model, state], outputs=[provider, api_key])
+    only_n_images.change(fn=update_only_n_images, inputs=[only_n_images, state], outputs=None)
+    provider.change(fn=update_provider, inputs=[provider, state], outputs=api_key)
+    api_key.change(fn=update_api_key, inputs=[api_key, state], outputs=None)
+    chatbot.clear(fn=clear_chat, inputs=[state], outputs=[chatbot])
+
+    # File upload event handlers
+    upload_button.click(
+        fn=handle_file_upload,
+        inputs=[file_upload, state],
+        outputs=[view_file_dropdown]
+    )
+    
+    # File viewing handlers
+    view_file_dropdown.change(
+        fn=view_file,
+        inputs=[view_file_dropdown, view_toggle],
+        outputs=[display_area]
+    )
+    
+    submit_button.click(process_input, [chat_input, state], [chatbot, view_file_dropdown])
+    stop_button.click(stop_app, [state], None)
+    
+    # Toggle view handler
+    view_toggle.change(
+        fn=toggle_view, 
+        inputs=[view_toggle, view_file_dropdown, state], 
+        outputs=[display_area, view_file_dropdown]
+    )
+    
+    # Refresh files handler
+    refresh_button.click(fn=refresh_files, inputs=[state], outputs=[view_file_dropdown])
+    
+    # Add JavaScript for auto-refresh instead of using demo.load()
+    js_refresh = """
+    function() {
+        // Auto-refresh files every 5 seconds
+        const refreshInterval = setInterval(function() {
+            // Find and click the refresh button
+            const refreshButtons = document.querySelectorAll('button');
+            for (const button of refreshButtons) {
+                if (button.textContent.includes('Refresh Files')) {
+                    button.click();
+                    break;
+                }
+            }
+        }, 5000);
+        
+        // Return a cleanup function
+        return () => clearInterval(refreshInterval);
+    }
+    """
+    
+    # Add the JavaScript to the page
+    gr.HTML("<script>(" + js_refresh + ")();</script>")
+    
+if __name__ == "__main__":
+    demo.launch(server_name="0.0.0.0", server_port=7888)

OmniParser/omnitool/gradio/app_streamlit.py

@@ -0,0 +1,470 @@
+"""
+Streamlit implementation of the OmniTool frontend.
+Usage: streamlit run app_streamlit.py -- --windows_host_url localhost:8006 --omniparser_server_url localhost:8000
+"""
+
+import os
+import io
+import shutil
+import mimetypes
+import argparse
+import base64
+from datetime import datetime
+from pathlib import Path
+from typing import cast
+from enum import StrEnum
+import streamlit as st
+from anthropic import APIResponse
+from anthropic.types import TextBlock
+from anthropic.types.beta import BetaMessage, BetaTextBlock, BetaToolUseBlock
+from anthropic.types.tool_use_block import ToolUseBlock
+import requests
+from requests.exceptions import RequestException
+
+from loop import (
+    APIProvider,
+    sampling_loop_sync,
+)
+from tools import ToolResult
+
+# Constants and configurations
+CONFIG_DIR = Path("~/.anthropic").expanduser()
+API_KEY_FILE = CONFIG_DIR / "api_key"
+UPLOAD_FOLDER = Path("./uploads").absolute()
+UPLOAD_FOLDER.mkdir(parents=True, exist_ok=True)
+
+class Sender(StrEnum):
+    USER = "user"
+    BOT = "assistant"
+    TOOL = "tool"
+
+def parse_arguments():
+    parser = argparse.ArgumentParser(description="Streamlit App")
+    parser.add_argument("--windows_host_url", type=str, default='localhost:8006')
+    parser.add_argument("--omniparser_server_url", type=str, default="localhost:8000")
+    parser.add_argument("--upload_folder", type=str, default="./uploads")
+    return parser.parse_known_args()[0]
+
+def initialize_session_state():
+    """Initialize session state variables"""
+    if "messages" not in st.session_state:
+        st.session_state.messages = []
+    if "model" not in st.session_state:
+        st.session_state.model = "omniparser + gpt-4o-orchestrated"
+    if "provider" not in st.session_state:
+        st.session_state.provider = "openai"
+    if "api_key" not in st.session_state:
+        st.session_state.api_key = os.getenv("OPENAI_API_KEY", "")
+    if "anthropic_api_key" not in st.session_state:
+        st.session_state.anthropic_api_key = os.getenv("ANTHROPIC_API_KEY", "")
+    if "only_n_most_recent_images" not in st.session_state:
+        st.session_state.only_n_most_recent_images = 2
+    if "responses" not in st.session_state:
+        st.session_state.responses = {}
+    if "tools" not in st.session_state:
+        st.session_state.tools = {}
+    if "uploaded_files" not in st.session_state:
+        st.session_state.uploaded_files = []
+    if "selected_file" not in st.session_state:
+        st.session_state.selected_file = "None"
+    if "stop" not in st.session_state:
+        st.session_state.stop = False
+
+def get_file_viewer_html(file_path=None, windows_host_url=None):
+    """Generate HTML to view a file based on its type"""
+    if not file_path:
+        # Return the VNC viewer iframe
+        return f'<iframe src="http://{windows_host_url}/vnc.html?view_only=1&autoconnect=1&resize=scale" width="100%" height="580" allow="fullscreen"></iframe>'
+    
+    file_path = Path(file_path)
+    if not file_path.exists():
+        return f'<div class="error-message">File not found: {file_path.name}</div>'
+    
+    mime_type, _ = mimetypes.guess_type(file_path)
+    file_type = mime_type.split('/')[0] if mime_type else 'unknown'
+    file_extension = file_path.suffix.lower()
+    
+    if file_type == 'image':
+        with open(file_path, "rb") as image_file:
+            encoded_string = base64.b64encode(image_file.read()).decode()
+            return f'<div class="file-viewer"><h3>{file_path.name}</h3><img src="data:{mime_type};base64,{encoded_string}" style="max-width:100%; max-height:500px;"></div>'
+    
+    elif file_extension in ['.txt', '.py', '.js', '.html', '.css', '.json', '.md', '.csv'] or file_type == 'text':
+        try:
+            content = file_path.read_text(errors='replace')
+            content = content.replace('&', '&amp;').replace('<', '&lt;').replace('>', '&gt;')
+            return f'<div class="file-viewer"><h3>{file_path.name}</h3><pre style="background-color: #f5f5f5; padding: 10px; border-radius: 5px; overflow: auto; max-height: 500px; white-space: pre-wrap;"><code>{content}</code></pre></div>'
+        except UnicodeDecodeError:
+            return f'<div class="error-message">Cannot display binary file: {file_path.name}</div>'
+    
+    else:
+        size_kb = file_path.stat().st_size / 1024
+        return f'<div class="file-viewer"><h3>{file_path.name}</h3><p>File type: {mime_type or "Unknown"}</p><p>Size: {size_kb:.2f} KB</p><p>This file type cannot be displayed in the browser.</p></div>'
+
+def handle_file_upload(uploaded_files):
+    """Handle file uploads and store them in the upload directory"""
+    if uploaded_files:
+        for file in uploaded_files:
+            file_path = UPLOAD_FOLDER / file.name
+            with open(file_path, "wb") as f:
+                f.write(file.getvalue())
+            if str(file_path) not in st.session_state.uploaded_files:
+                st.session_state.uploaded_files.append(str(file_path))
+
+def _api_response_callback(response: APIResponse[BetaMessage]):
+    response_id = datetime.now().isoformat()
+    st.session_state.responses[response_id] = response
+
+def _tool_output_callback(tool_output: ToolResult, tool_id: str):
+    st.session_state.tools[tool_id] = tool_output
+
+def chatbot_output_callback(message, hide_images=False):
+    def _render_message(message: str | BetaTextBlock | BetaToolUseBlock | ToolResult, hide_images=False):
+        if isinstance(message, str):
+            return message
+        
+        is_tool_result = not isinstance(message, str) and (
+            isinstance(message, ToolResult)
+            or message.__class__.__name__ == "ToolResult"
+        )
+        
+        if is_tool_result:
+            message = cast(ToolResult, message)
+            if message.output:
+                return message.output
+            if message.error:
+                return f"Error: {message.error}"
+            if message.base64_image and not hide_images:
+                return f'<img src="data:image/png;base64,{message.base64_image}">'
+        
+        elif isinstance(message, (BetaTextBlock, TextBlock)):
+            return f"Next step Reasoning: {message.text}"
+        
+        elif isinstance(message, (BetaToolUseBlock, ToolUseBlock)):
+            return None
+        
+        return message
+
+    rendered_message = _render_message(message, hide_images)
+    if rendered_message:
+        st.session_state.messages.append({"role": "assistant", "content": rendered_message})
+
+def main():
+    args = parse_arguments()
+    initialize_session_state()
+
+    # Page configuration
+    st.set_page_config(
+        page_title="OmniTool",
+        page_icon="🤖",
+        layout="wide"
+    )
+
+    # Custom CSS
+    st.markdown("""
+        <style>
+        .stApp {
+            max-width: 100%;
+            padding: 1rem;
+        }
+        .chat-container {
+            height: calc(100vh - 200px);
+            overflow-y: auto;
+            position: relative;
+        }
+        .viewer-container {
+            height: calc(100vh - 200px);
+            overflow-y: auto;
+        }
+        .chat-input-container {
+            display: flex;
+            align-items: flex-end;
+        }
+        .icon-button {
+            border: none;
+            background: none;
+            cursor: pointer;
+            padding: 0.5rem;
+            border-radius: 50%;
+            transition: background-color 0.3s;
+        }
+        .icon-button:hover {
+            background-color: #f0f0f0;
+        }
+        .stButton button {
+            border-radius: 50%;
+            width: 40px;
+            height: 40px;
+            padding: 0;
+            display: flex;
+            align-items: center;
+            justify-content: center;
+            box-shadow: 0 1px 3px rgba(0,0,0,0.1);
+        }
+        /* Custom button styles */
+        .send-btn {
+            background-color: #000 !important;
+            color: white !important;
+        }
+        .stop-btn {
+            background-color: #f8f9fa !important;
+            color: #d9534f !important;
+            border: 1px solid #d9534f !important;
+        }
+        .upload-btn {
+            background-color: #f8f9fa !important;
+            color: #0275d8 !important;
+            border: 1px solid #0275d8 !important;
+        }
+        /* Hide the default button styling */
+        div[data-testid="stHorizontalBlock"] button[kind="secondary"] {
+            background-color: transparent;
+            border: none;
+        }
+        /* Share button positioning */
+        .share-button-container {
+            position: absolute;
+            top: 0;
+            right: 0;
+            z-index: 100;
+        }
+        /* Chat header with title and share button */
+        .chat-header {
+            display: flex;
+            justify-content: space-between;
+            align-items: center;
+            margin-bottom: 10px;
+        }
+        /* Input placeholder styling */
+        .stTextInput input::placeholder {
+            color: #6c757d;
+            font-style: italic;
+        }
+        </style>
+    """, unsafe_allow_html=True)
+
+    # Header
+    st.title("OmniTool")
+
+    # Sidebar with settings
+    with st.sidebar:
+        st.header("Settings")
+        
+        # Model selection
+        model = st.selectbox(
+            "Model",
+            ["omniparser + gpt-4o", "omniparser + o1", "omniparser + o3-mini", 
+             "omniparser + R1", "omniparser + qwen2.5vl", "claude-3-5-sonnet-20241022",
+             "omniparser + gpt-4o-orchestrated", "omniparser + o1-orchestrated",
+             "omniparser + o3-mini-orchestrated", "omniparser + R1-orchestrated",
+             "omniparser + qwen2.5vl-orchestrated"],
+            index=6
+        )
+        st.session_state.model = model
+
+        # API settings
+        api_key = st.text_input("API Key", value=st.session_state.api_key, type="password")
+        st.session_state.api_key = api_key
+
+        # Image settings
+        n_images = st.slider("N most recent screenshots", 0, 10, 2)
+        st.session_state.only_n_most_recent_images = n_images
+
+        # File viewer selection
+        file_options = ["None"]
+        if st.session_state.uploaded_files:
+            file_options.extend([Path(f).name for f in st.session_state.uploaded_files])
+        
+        selected_file = st.selectbox(
+            "View File",
+            options=file_options,
+            format_func=lambda x: x
+        )
+        st.session_state.selected_file = selected_file
+        
+        view_mode = st.radio("Display Mode", ["OmniTool Computer", "File Viewer"])
+
+    # Main content area with two columns
+    col1, col2 = st.columns([2, 3])
+
+    # Chat interface (left column)
+    with col1:
+        # Chat header with title and share button
+        col_header_1, col_header_2 = st.columns([3, 1])
+        with col_header_1:
+            st.markdown("### Chat")
+        with col_header_2:
+            share_button = st.button("📤 Share", key="share_btn", help="Share conversation")
+            # Apply custom styling with HTML
+            st.markdown("""
+                <style>
+                button[data-testid="share_btn"] {
+                    background-color: #f8f9fa !important;
+                    color: #0275d8 !important;
+                    border: 1px solid #0275d8 !important;
+                    border-radius: 4px !important;
+                    width: auto !important;
+                    height: auto !important;
+                    padding: 2px 8px !important;
+                    font-size: 0.8rem !important;
+                }
+                </style>
+            """, unsafe_allow_html=True)
+        
+        # Share functionality
+        if share_button:
+            # Create a shareable text of the conversation
+            conversation_text = ""
+            for message in st.session_state.messages:
+                if message["role"] == "user":
+                    conversation_text += f"User: {message['content']}\n\n"
+                else:
+                    conversation_text += f"Assistant: {message['content']}\n\n"
+            
+            # Create a download link
+            timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+            st.download_button(
+                label="Download Conversation",
+                data=conversation_text,
+                file_name=f"omnitool_conversation_{timestamp}.txt",
+                mime="text/plain",
+                key="download_conversation"
+            )
+        
+        # Display chat messages
+        chat_container = st.container(height=450)
+        with chat_container:
+            for message in st.session_state.messages:
+                if message["role"] == "user":
+                    st.markdown(f"**You:** {message['content']}")
+                else:
+                    st.markdown(f"**Assistant:** {message['content']}", unsafe_allow_html=True)
+
+        # Chat input and buttons
+        user_input = st.text_input(
+            "Type your message:", 
+            key="user_input", 
+            label_visibility="collapsed",
+            placeholder="Send message to OmniTool..."
+        )
+        
+        # Button row with icons
+        col1_1, col1_2, col1_3, col1_4 = st.columns([6, 1, 1, 1])
+        
+        with col1_2:
+            # Send button with icon - using arrow up icon
+            send_button = st.button("⬆️", help="Send message", key="send_btn")
+            # Apply custom styling with HTML
+            st.markdown("""
+                <style>
+                button[data-testid="send_btn"] {
+                    background-color: black !important;
+                    color: white !important;
+                }
+                </style>
+            """, unsafe_allow_html=True)
+        
+        with col1_3:
+            # Stop button with icon
+            stop_button = st.button("🛑", help="Stop processing", key="stop_btn")
+            # Apply custom styling with HTML
+            st.markdown("""
+                <style>
+                button[data-testid="stop_btn"] {
+                    background-color: #f8f9fa !important;
+                    color: #d9534f !important;
+                    border: 1px solid #d9534f !important;
+                }
+                </style>
+            """, unsafe_allow_html=True)
+            
+        with col1_4:
+            # File upload button with icon
+            upload_button = st.button("📎", help="Upload files", key="upload_btn")
+            # Apply custom styling with HTML
+            st.markdown("""
+                <style>
+                button[data-testid="upload_btn"] {
+                    background-color: #f8f9fa !important;
+                    color: #0275d8 !important;
+                    border: 1px solid #0275d8 !important;
+                }
+                </style>
+            """, unsafe_allow_html=True)
+        
+        # File upload area (hidden by default, shown when upload button is clicked)
+        if upload_button:
+            uploaded_files = st.file_uploader("Upload Files", accept_multiple_files=True, label_visibility="collapsed")
+            if uploaded_files:
+                handle_file_upload(uploaded_files)
+                st.success(f"Uploaded {len(uploaded_files)} file(s)")
+                # Update file options
+                file_options = ["None"]
+                if st.session_state.uploaded_files:
+                    file_options.extend([Path(f).name for f in st.session_state.uploaded_files])
+                st.rerun()
+        
+        # Process send button click
+        if send_button and user_input:
+            # Add user message to state
+            st.session_state.messages.append({"role": "user", "content": user_input})
+            
+            # Process the message through sampling_loop_sync
+            for loop_msg in sampling_loop_sync(
+                model=st.session_state.model,
+                provider=st.session_state.provider,
+                messages=[{"role": "user", "content": [TextBlock(type="text", text=msg["content"])]} for msg in st.session_state.messages],
+                output_callback=chatbot_output_callback,
+                tool_output_callback=_tool_output_callback,
+                api_response_callback=_api_response_callback,
+                api_key=st.session_state.api_key,
+                only_n_most_recent_images=st.session_state.only_n_most_recent_images,
+                max_tokens=16384,
+                omniparser_url=args.omniparser_server_url,
+                save_folder=str(UPLOAD_FOLDER)
+            ):
+                if loop_msg is None or st.session_state.stop:
+                    break
+                st.rerun()
+        
+        # Process stop button click
+        if stop_button:
+            st.session_state.stop = True
+            st.info("Processing stopped")
+
+    # Viewer interface (right column)
+    with col2:
+        st.markdown("### Display")
+        if view_mode == "OmniTool Computer":
+            viewer_html = get_file_viewer_html(windows_host_url=args.windows_host_url)
+            st.components.v1.html(
+                viewer_html,
+                height=600,
+                scrolling=True
+            )
+        else:  # File Viewer mode
+            if st.session_state.selected_file and st.session_state.selected_file != "None":
+                file_path = next((f for f in st.session_state.uploaded_files 
+                                if Path(f).name == st.session_state.selected_file), None)
+                if file_path:
+                    viewer_html = get_file_viewer_html(file_path=file_path)
+                    st.components.v1.html(
+                        viewer_html,
+                        height=600,
+                        scrolling=True
+                    )
+                else:
+                    st.error(f"Could not find file: {st.session_state.selected_file}")
+            else:
+                st.info("Please select a file to view from the sidebar.")
+
+        # Debug information (temporary)
+        with st.expander("Debug Info"):
+            st.write("View Mode:", view_mode)
+            st.write("Selected File:", st.session_state.selected_file)
+            st.write("Available Files:", st.session_state.uploaded_files)
+            if view_mode == "File Viewer" and st.session_state.selected_file != "None":
+                st.write("File Path:", file_path if 'file_path' in locals() else "Not found")
+
+if __name__ == "__main__":
+    main()

OmniParser/omnitool/gradio/executor/anthropic_executor.py

@@ -0,0 +1,132 @@
+import asyncio
+from typing import Any, Dict, cast
+from collections.abc import Callable
+from anthropic.types.beta import (
+    BetaContentBlock,
+    BetaContentBlockParam,
+    BetaImageBlockParam,
+    BetaMessage,
+    BetaMessageParam,
+    BetaTextBlockParam,
+    BetaToolResultBlockParam,
+)
+from anthropic.types import TextBlock
+from anthropic.types.beta import BetaMessage, BetaTextBlock, BetaToolUseBlock
+from tools import ComputerTool, ToolCollection, ToolResult
+
+
+class AnthropicExecutor:
+    def __init__(
+        self, 
+        output_callback: Callable[[BetaContentBlockParam], None], 
+        tool_output_callback: Callable[[Any, str], None],
+    ):
+        self.tool_collection = ToolCollection(
+            ComputerTool()
+        )
+        self.output_callback = output_callback
+        self.tool_output_callback = tool_output_callback
+
+    def __call__(self, response: BetaMessage, messages: list[BetaMessageParam]):
+        new_message = {
+            "role": "assistant",
+            "content": cast(list[BetaContentBlockParam], response.content),
+        }
+        if new_message not in messages:
+            messages.append(new_message)
+        else:
+            print("new_message already in messages, there are duplicates.")
+        
+        tool_result_content: list[BetaToolResultBlockParam] = []
+        for content_block in cast(list[BetaContentBlock], response.content):
+            self.output_callback(content_block, sender="bot")
+            # Execute the tool
+            if content_block.type == "tool_use":
+                # Run the asynchronous tool execution in a synchronous context
+                result = asyncio.run(self.tool_collection.run(
+                    name=content_block.name,
+                    tool_input=cast(dict[str, Any], content_block.input),
+                ))
+                
+                self.output_callback(result, sender="bot")
+                
+                tool_result_content.append(
+                    _make_api_tool_result(result, content_block.id)
+                )
+                # self.tool_output_callback(result, content_block.id)
+
+            # Craft messages based on the content_block
+            # Note: to display the messages in the gradio, you should organize the messages in the following way (user message, bot message)
+            
+            display_messages = _message_display_callback(messages)
+            # display_messages = []
+            
+            # Send the messages to the gradio
+            for user_msg, bot_msg in display_messages:
+                # yield [user_msg, bot_msg], tool_result_content
+                yield [None, None], tool_result_content
+
+        if not tool_result_content:
+            return messages
+        
+        return tool_result_content
+
+def _message_display_callback(messages):
+    display_messages = []
+    for msg in messages:
+        try:
+            if isinstance(msg["content"][0], TextBlock):
+                display_messages.append((msg["content"][0].text, None))  # User message
+            elif isinstance(msg["content"][0], BetaTextBlock):
+                display_messages.append((None, msg["content"][0].text))  # Bot message
+            elif isinstance(msg["content"][0], BetaToolUseBlock):
+                display_messages.append((None, f"Tool Use: {msg['content'][0].name}\nInput: {msg['content'][0].input}"))  # Bot message
+            elif isinstance(msg["content"][0], Dict) and msg["content"][0]["content"][-1]["type"] == "image":
+                display_messages.append((None, f'<img src="data:image/png;base64,{msg["content"][0]["content"][-1]["source"]["data"]}">'))  # Bot message
+            else:
+                print(msg["content"][0])
+        except Exception as e:
+            print("error", e)
+            pass
+    return display_messages
+
+def _make_api_tool_result(
+    result: ToolResult, tool_use_id: str
+) -> BetaToolResultBlockParam:
+    """Convert an agent ToolResult to an API ToolResultBlockParam."""
+    tool_result_content: list[BetaTextBlockParam | BetaImageBlockParam] | str = []
+    is_error = False
+    if result.error:
+        is_error = True
+        tool_result_content = _maybe_prepend_system_tool_result(result, result.error)
+    else:
+        if result.output:
+            tool_result_content.append(
+                {
+                    "type": "text",
+                    "text": _maybe_prepend_system_tool_result(result, result.output),
+                }
+            )
+        if result.base64_image:
+            tool_result_content.append(
+                {
+                    "type": "image",
+                    "source": {
+                        "type": "base64",
+                        "media_type": "image/png",
+                        "data": result.base64_image,
+                    },
+                }
+            )
+    return {
+        "type": "tool_result",
+        "content": tool_result_content,
+        "tool_use_id": tool_use_id,
+        "is_error": is_error,
+    }
+
+
+def _maybe_prepend_system_tool_result(result: ToolResult, result_text: str):
+    if result.system:
+        result_text = f"<system>{result.system}</system>\n{result_text}"
+    return result_text

OmniParser/omnitool/gradio/loop.py

@@ -0,0 +1,127 @@
+"""
+Agentic sampling loop that calls the Anthropic API and local implenmentation of anthropic-defined computer use tools.
+"""
+from collections.abc import Callable
+from enum import StrEnum
+
+from anthropic import APIResponse
+from anthropic.types import (
+    TextBlock,
+)
+from anthropic.types.beta import (
+    BetaContentBlock,
+    BetaMessage,
+    BetaMessageParam
+)
+from tools import ToolResult
+
+from agent.llm_utils.omniparserclient import OmniParserClient
+from agent.anthropic_agent import AnthropicActor
+from agent.vlm_agent import VLMAgent
+from agent.vlm_agent_with_orchestrator import VLMOrchestratedAgent
+from executor.anthropic_executor import AnthropicExecutor
+
+BETA_FLAG = "computer-use-2024-10-22"
+
+class APIProvider(StrEnum):
+    ANTHROPIC = "anthropic"
+    BEDROCK = "bedrock"
+    VERTEX = "vertex"
+    OPENAI = "openai"
+
+
+PROVIDER_TO_DEFAULT_MODEL_NAME: dict[APIProvider, str] = {
+    APIProvider.ANTHROPIC: "claude-3-5-sonnet-20241022",
+    APIProvider.BEDROCK: "anthropic.claude-3-5-sonnet-20241022-v2:0",
+    APIProvider.VERTEX: "claude-3-5-sonnet-v2@20241022",
+    APIProvider.OPENAI: "gpt-4o",
+}
+
+def sampling_loop_sync(
+    *,
+    model: str,
+    provider: APIProvider | None,
+    messages: list[BetaMessageParam],
+    output_callback: Callable[[BetaContentBlock], None],
+    tool_output_callback: Callable[[ToolResult, str], None],
+    api_response_callback: Callable[[APIResponse[BetaMessage]], None],
+    api_key: str,
+    only_n_most_recent_images: int | None = 2,
+    max_tokens: int = 4096,
+    omniparser_url: str,
+    save_folder: str = "./uploads"
+):
+    """
+    Synchronous agentic sampling loop for the assistant/tool interaction of computer use.
+    """
+    print('in sampling_loop_sync, model:', model)
+    omniparser_client = OmniParserClient(url=f"http://{omniparser_url}/parse/")
+    if model == "claude-3-5-sonnet-20241022":
+        # Register Actor and Executor
+        actor = AnthropicActor(
+            model=model, 
+            provider=provider,
+            api_key=api_key, 
+            api_response_callback=api_response_callback,
+            max_tokens=max_tokens,
+            only_n_most_recent_images=only_n_most_recent_images
+        )
+    elif model in set(["omniparser + gpt-4o", "omniparser + o1", "omniparser + o3-mini", "omniparser + R1", "omniparser + qwen2.5vl"]):
+        actor = VLMAgent(
+            model=model,
+            provider=provider,
+            api_key=api_key,
+            api_response_callback=api_response_callback,
+            output_callback=output_callback,
+            max_tokens=max_tokens,
+            only_n_most_recent_images=only_n_most_recent_images
+        )
+    elif model in set(["omniparser + gpt-4o-orchestrated", "omniparser + o1-orchestrated", "omniparser + o3-mini-orchestrated", "omniparser + R1-orchestrated", "omniparser + qwen2.5vl-orchestrated"]):
+        actor = VLMOrchestratedAgent(
+            model=model,
+            provider=provider,
+            api_key=api_key,
+            api_response_callback=api_response_callback,
+            output_callback=output_callback,
+            max_tokens=max_tokens,
+            only_n_most_recent_images=only_n_most_recent_images,
+            save_folder=save_folder
+        )
+    else:
+        raise ValueError(f"Model {model} not supported")
+    executor = AnthropicExecutor(
+        output_callback=output_callback,
+        tool_output_callback=tool_output_callback,
+    )
+    print(f"Model Inited: {model}, Provider: {provider}")
+    
+    tool_result_content = None
+    
+    print(f"Start the message loop. User messages: {messages}")
+    
+    if model == "claude-3-5-sonnet-20241022": # Anthropic loop
+        while True:
+            parsed_screen = omniparser_client() # parsed_screen: {"som_image_base64": dino_labled_img, "parsed_content_list": parsed_content_list, "screen_info"}
+            screen_info_block = TextBlock(text='Below is the structured accessibility information of the current UI screen, which includes text and icons you can operate on, take these information into account when you are making the prediction for the next action. Note you will still need to take screenshot to get the image: \n' + parsed_screen['screen_info'], type='text')
+            screen_info_dict = {"role": "user", "content": [screen_info_block]}
+            messages.append(screen_info_dict)
+            tools_use_needed = actor(messages=messages)
+
+            for message, tool_result_content in executor(tools_use_needed, messages):
+                yield message
+        
+            if not tool_result_content:
+                return messages
+
+            messages.append({"content": tool_result_content, "role": "user"})
+    
+    elif model in set(["omniparser + gpt-4o", "omniparser + o1", "omniparser + o3-mini", "omniparser + R1", "omniparser + qwen2.5vl", "omniparser + gpt-4o-orchestrated", "omniparser + o1-orchestrated", "omniparser + o3-mini-orchestrated", "omniparser + R1-orchestrated", "omniparser + qwen2.5vl-orchestrated"]):
+        while True:
+            parsed_screen = omniparser_client()
+            tools_use_needed, vlm_response_json = actor(messages=messages, parsed_screen=parsed_screen)
+
+            for message, tool_result_content in executor(tools_use_needed, messages):
+                yield message
+        
+            if not tool_result_content:
+                return messages

OmniParser/omnitool/gradio/tools/__init__.py

@@ -0,0 +1,11 @@
+from .base import ToolResult
+from .collection import ToolCollection
+from .computer import ComputerTool
+from .screen_capture import get_screenshot
+
+__ALL__ = [
+    ComputerTool,
+    ToolCollection,
+    ToolResult,
+    get_screenshot,
+]

OmniParser/omnitool/gradio/tools/base.py

@@ -0,0 +1,65 @@
+from abc import ABCMeta, abstractmethod
+from dataclasses import dataclass, fields, replace
+from typing import Any
+
+from anthropic.types.beta import BetaToolUnionParam
+
+
+class BaseAnthropicTool(metaclass=ABCMeta):
+    """Abstract base class for Anthropic-defined tools."""
+
+    @abstractmethod
+    def __call__(self, **kwargs) -> Any:
+        """Executes the tool with the given arguments."""
+        ...
+
+    @abstractmethod
+    def to_params(
+        self,
+    ) -> BetaToolUnionParam:
+        raise NotImplementedError
+
+
+@dataclass(kw_only=True, frozen=True)
+class ToolResult:
+    """Represents the result of a tool execution."""
+
+    output: str | None = None
+    error: str | None = None
+    base64_image: str | None = None
+    system: str | None = None
+
+    def __bool__(self):
+        return any(getattr(self, field.name) for field in fields(self))
+
+    def __add__(self, other: "ToolResult"):
+        def combine_fields(
+            field: str | None, other_field: str | None, concatenate: bool = True
+        ):
+            if field and other_field:
+                if concatenate:
+                    return field + other_field
+                raise ValueError("Cannot combine tool results")
+            return field or other_field
+
+        return ToolResult(
+            output=combine_fields(self.output, other.output),
+            error=combine_fields(self.error, other.error),
+            base64_image=combine_fields(self.base64_image, other.base64_image, False),
+            system=combine_fields(self.system, other.system),
+        )
+
+    def replace(self, **kwargs):
+        """Returns a new ToolResult with the given fields replaced."""
+        return replace(self, **kwargs)
+
+
+class ToolFailure(ToolResult):
+    """A ToolResult that represents a failure."""
+
+
+class ToolError(Exception):
+    """Raised when a tool encounters an error."""
+
+    def __init__(self, message):
+        self.message = message

OmniParser/omnitool/gradio/tools/collection.py

@@ -0,0 +1,34 @@
+"""Collection classes for managing multiple tools."""
+
+from typing import Any
+
+from anthropic.types.beta import BetaToolUnionParam
+
+from .base import (
+    BaseAnthropicTool,
+    ToolError,
+    ToolFailure,
+    ToolResult,
+)
+
+
+class ToolCollection:
+    """A collection of anthropic-defined tools."""
+
+    def __init__(self, *tools: BaseAnthropicTool):
+        self.tools = tools
+        self.tool_map = {tool.to_params()["name"]: tool for tool in tools}
+
+    def to_params(
+        self,
+    ) -> list[BetaToolUnionParam]:
+        return [tool.to_params() for tool in self.tools]
+
+    async def run(self, *, name: str, tool_input: dict[str, Any]) -> ToolResult:
+        tool = self.tool_map.get(name)
+        if not tool:
+            return ToolFailure(error=f"Tool {name} is invalid")
+        try:
+            return await tool(**tool_input)
+        except ToolError as e:
+            return ToolFailure(error=e.message)

OmniParser/omnitool/gradio/tools/computer.py

@@ -0,0 +1,329 @@
+import base64
+import time
+from enum import StrEnum
+from typing import Literal, TypedDict
+
+from PIL import Image
+
+from anthropic.types.beta import BetaToolComputerUse20241022Param
+
+from .base import BaseAnthropicTool, ToolError, ToolResult
+from .screen_capture import get_screenshot
+import requests
+import re
+
+OUTPUT_DIR = "./tmp/outputs"
+
+TYPING_DELAY_MS = 12
+TYPING_GROUP_SIZE = 50
+
+Action = Literal[
+    "key",
+    "type",
+    "mouse_move",
+    "left_click",
+    "left_click_drag",
+    "right_click",
+    "middle_click",
+    "double_click",
+    "screenshot",
+    "cursor_position",
+    "hover",
+    "wait"
+]
+
+
+class Resolution(TypedDict):
+    width: int
+    height: int
+
+
+MAX_SCALING_TARGETS: dict[str, Resolution] = {
+    "XGA": Resolution(width=1024, height=768),  # 4:3
+    "WXGA": Resolution(width=1280, height=800),  # 16:10
+    "FWXGA": Resolution(width=1366, height=768),  # ~16:9
+}
+
+
+class ScalingSource(StrEnum):
+    COMPUTER = "computer"
+    API = "api"
+
+
+class ComputerToolOptions(TypedDict):
+    display_height_px: int
+    display_width_px: int
+    display_number: int | None
+
+
+def chunks(s: str, chunk_size: int) -> list[str]:
+    return [s[i : i + chunk_size] for i in range(0, len(s), chunk_size)]
+
+class ComputerTool(BaseAnthropicTool):
+    """
+    A tool that allows the agent to interact with the screen, keyboard, and mouse of the current computer.
+    Adapted for Windows using 'pyautogui'.
+    """
+
+    name: Literal["computer"] = "computer"
+    api_type: Literal["computer_20241022"] = "computer_20241022"
+    width: int
+    height: int
+    display_num: int | None
+
+    _screenshot_delay = 2.0
+    _scaling_enabled = True
+
+    @property
+    def options(self) -> ComputerToolOptions:
+        width, height = self.scale_coordinates(
+            ScalingSource.COMPUTER, self.width, self.height
+        )
+        return {
+            "display_width_px": width,
+            "display_height_px": height,
+            "display_number": self.display_num,
+        }
+
+    def to_params(self) -> BetaToolComputerUse20241022Param:
+        return {"name": self.name, "type": self.api_type, **self.options}
+
+    def __init__(self, is_scaling: bool = False):
+        super().__init__()
+
+        # Get screen width and height using Windows command
+        self.display_num = None
+        self.offset_x = 0
+        self.offset_y = 0
+        self.is_scaling = is_scaling
+        self.width, self.height = self.get_screen_size()
+        print(f"screen size: {self.width}, {self.height}")
+
+        self.key_conversion = {"Page_Down": "pagedown",
+                               "Page_Up": "pageup",
+                               "Super_L": "win",
+                               "Escape": "esc"}
+
+
+    async def __call__(
+        self,
+        *,
+        action: Action,
+        text: str | None = None,
+        coordinate: tuple[int, int] | None = None,
+        **kwargs,
+    ):
+        print(f"action: {action}, text: {text}, coordinate: {coordinate}, is_scaling: {self.is_scaling}")
+        if action in ("mouse_move", "left_click_drag"):
+            if coordinate is None:
+                raise ToolError(f"coordinate is required for {action}")
+            if text is not None:
+                raise ToolError(f"text is not accepted for {action}")
+            if not isinstance(coordinate, (list, tuple)) or len(coordinate) != 2:
+                raise ToolError(f"{coordinate} must be a tuple of length 2")
+            # if not all(isinstance(i, int) and i >= 0 for i in coordinate):
+            if not all(isinstance(i, int) for i in coordinate):
+                raise ToolError(f"{coordinate} must be a tuple of non-negative ints")
+            
+            if self.is_scaling:
+                x, y = self.scale_coordinates(
+                    ScalingSource.API, coordinate[0], coordinate[1]
+                )
+            else:
+                x, y = coordinate
+
+            # print(f"scaled_coordinates: {x}, {y}")
+            # print(f"offset: {self.offset_x}, {self.offset_y}")
+            
+            # x += self.offset_x # TODO - check if this is needed
+            # y += self.offset_y
+
+            print(f"mouse move to {x}, {y}")
+            
+            if action == "mouse_move":
+                self.send_to_vm(f"pyautogui.moveTo({x}, {y})")
+                return ToolResult(output=f"Moved mouse to ({x}, {y})")
+            elif action == "left_click_drag":
+                current_x, current_y = self.send_to_vm("pyautogui.position()")
+                self.send_to_vm(f"pyautogui.dragTo({x}, {y}, duration=0.5)")
+                return ToolResult(output=f"Dragged mouse from ({current_x}, {current_y}) to ({x}, {y})")
+
+        if action in ("key", "type"):
+            if text is None:
+                raise ToolError(f"text is required for {action}")
+            if coordinate is not None:
+                raise ToolError(f"coordinate is not accepted for {action}")
+            if not isinstance(text, str):
+                raise ToolError(output=f"{text} must be a string")
+
+            if action == "key":
+                # Handle key combinations
+                keys = text.split('+')
+                for key in keys:
+                    key = self.key_conversion.get(key.strip(), key.strip())
+                    key = key.lower()
+                    self.send_to_vm(f"pyautogui.keyDown('{key}')")  # Press down each key
+                for key in reversed(keys):
+                    key = self.key_conversion.get(key.strip(), key.strip())
+                    key = key.lower()
+                    self.send_to_vm(f"pyautogui.keyUp('{key}')")    # Release each key in reverse order
+                return ToolResult(output=f"Pressed keys: {text}")
+            
+            elif action == "type":
+                # default click before type TODO: check if this is needed
+                self.send_to_vm("pyautogui.click()")
+                self.send_to_vm(f"pyautogui.typewrite('{text}', interval={TYPING_DELAY_MS / 1000})")
+                self.send_to_vm("pyautogui.press('enter')")
+                screenshot_base64 = (await self.screenshot()).base64_image
+                return ToolResult(output=text, base64_image=screenshot_base64)
+
+        if action in (
+            "left_click",
+            "right_click",
+            "double_click",
+            "middle_click",
+            "screenshot",
+            "cursor_position",
+            "left_press",
+        ):
+            if text is not None:
+                raise ToolError(f"text is not accepted for {action}")
+            if coordinate is not None:
+                raise ToolError(f"coordinate is not accepted for {action}")
+
+            if action == "screenshot":
+                return await self.screenshot()
+            elif action == "cursor_position":
+                x, y = self.send_to_vm("pyautogui.position()")
+                x, y = self.scale_coordinates(ScalingSource.COMPUTER, x, y)
+                return ToolResult(output=f"X={x},Y={y}")
+            else:
+                if action == "left_click":
+                    self.send_to_vm("pyautogui.click()")
+                elif action == "right_click":
+                    self.send_to_vm("pyautogui.rightClick()")
+                elif action == "middle_click":
+                    self.send_to_vm("pyautogui.middleClick()")
+                elif action == "double_click":
+                    self.send_to_vm("pyautogui.doubleClick()")
+                elif action == "left_press":
+                    self.send_to_vm("pyautogui.mouseDown()")
+                    time.sleep(1)
+                    self.send_to_vm("pyautogui.mouseUp()")
+                return ToolResult(output=f"Performed {action}")
+        if action in ("scroll_up", "scroll_down"):
+            if action == "scroll_up":
+                self.send_to_vm("pyautogui.scroll(100)")
+            elif action == "scroll_down":
+                self.send_to_vm("pyautogui.scroll(-100)")
+            return ToolResult(output=f"Performed {action}")
+        if action == "hover":
+            return ToolResult(output=f"Performed {action}")
+        if action == "wait":
+            time.sleep(1)
+            return ToolResult(output=f"Performed {action}")
+        raise ToolError(f"Invalid action: {action}")
+
+    def send_to_vm(self, action: str):
+        """
+        Executes a python command on the server. Only return tuple of x,y when action is "pyautogui.position()"
+        """
+        prefix = "import pyautogui; pyautogui.FAILSAFE = False;"
+        command_list = ["python", "-c", f"{prefix} {action}"]
+        parse = action == "pyautogui.position()"
+        if parse:
+            command_list[-1] = f"{prefix} print({action})"
+
+        try:
+            print(f"sending to vm: {command_list}")
+            response = requests.post(
+                f"http://localhost:5000/execute", 
+                headers={'Content-Type': 'application/json'},
+                json={"command": command_list},
+                timeout=90
+            )
+            time.sleep(0.7) # avoid async error as actions take time to complete
+            print(f"action executed")
+            if response.status_code != 200:
+                raise ToolError(f"Failed to execute command. Status code: {response.status_code}")
+            if parse:
+                output = response.json()['output'].strip()
+                match = re.search(r'Point\(x=(\d+),\s*y=(\d+)\)', output)
+                if not match:
+                    raise ToolError(f"Could not parse coordinates from output: {output}")
+                x, y = map(int, match.groups())
+                return x, y
+        except requests.exceptions.RequestException as e:
+            raise ToolError(f"An error occurred while trying to execute the command: {str(e)}")
+
+    async def screenshot(self):
+        if not hasattr(self, 'target_dimension'):
+            screenshot = self.padding_image(screenshot)
+            self.target_dimension = MAX_SCALING_TARGETS["WXGA"]
+        width, height = self.target_dimension["width"], self.target_dimension["height"]
+        screenshot, path = get_screenshot(resize=True, target_width=width, target_height=height)
+        time.sleep(0.7) # avoid async error as actions take time to complete
+        return ToolResult(base64_image=base64.b64encode(path.read_bytes()).decode())
+
+    def padding_image(self, screenshot):
+        """Pad the screenshot to 16:10 aspect ratio, when the aspect ratio is not 16:10."""
+        _, height = screenshot.size
+        new_width = height * 16 // 10
+
+        padding_image = Image.new("RGB", (new_width, height), (255, 255, 255))
+        # padding to top left
+        padding_image.paste(screenshot, (0, 0))
+        return padding_image
+
+    def scale_coordinates(self, source: ScalingSource, x: int, y: int):
+        """Scale coordinates to a target maximum resolution."""
+        if not self._scaling_enabled:
+            return x, y
+        ratio = self.width / self.height
+        target_dimension = None
+
+        for target_name, dimension in MAX_SCALING_TARGETS.items():
+            # allow some error in the aspect ratio - not ratios are exactly 16:9
+            if abs(dimension["width"] / dimension["height"] - ratio) < 0.02:
+                if dimension["width"] < self.width:
+                    target_dimension = dimension
+                    self.target_dimension = target_dimension
+                    # print(f"target_dimension: {target_dimension}")
+                break
+
+        if target_dimension is None:
+            # TODO: currently we force the target to be WXGA (16:10), when it cannot find a match
+            target_dimension = MAX_SCALING_TARGETS["WXGA"]
+            self.target_dimension = MAX_SCALING_TARGETS["WXGA"]
+
+        # should be less than 1
+        x_scaling_factor = target_dimension["width"] / self.width
+        y_scaling_factor = target_dimension["height"] / self.height
+        if source == ScalingSource.API:
+            if x > self.width or y > self.height:
+                raise ToolError(f"Coordinates {x}, {y} are out of bounds")
+            # scale up
+            return round(x / x_scaling_factor), round(y / y_scaling_factor)
+        # scale down
+        return round(x * x_scaling_factor), round(y * y_scaling_factor)
+
+    def get_screen_size(self):
+        """Return width and height of the screen"""
+        try:
+            response = requests.post(
+                f"http://localhost:5000/execute",
+                headers={'Content-Type': 'application/json'},
+                json={"command": ["python", "-c", "import pyautogui; print(pyautogui.size())"]},
+                timeout=90
+            )
+            if response.status_code != 200:
+                raise ToolError(f"Failed to get screen size. Status code: {response.status_code}")
+            
+            output = response.json()['output'].strip()
+            match = re.search(r'Size\(width=(\d+),\s*height=(\d+)\)', output)
+            if not match:
+                raise ToolError(f"Could not parse screen size from output: {output}")
+            width, height = map(int, match.groups())
+            return width, height
+        except requests.exceptions.RequestException as e:
+            raise ToolError(f"An error occurred while trying to get screen size: {str(e)}")