English β’ δΈζ β’ ζ₯ζ¬θͺ β’ νκ΅μ΄
Current language agent frameworks aim to facilitate the construction of proof-of-concept language agents while neglecting the non-expert user access to agents and paying little attention to application-level designs. We built OpenAgents, an open platform for using and hosting language agents in the wild of everyday life.
We have now implemented three agents in OpenAgents, and we host them on demo for free use!
- Data Agent for data analysis with Python/SQL and data tools;
- Plugins Agent with 200+ daily tools;
- Web Agent for autonomous web browsing.
OpenAgents can analyze data, call plugins, control your browser as ChatGPT Plus, but with OPEN Code for
- Easy deployment
- Full stack
- Chat Web UI
- Agent methods
- β¦
OpenAgents enables general users to interact with agent functionalities through a web UI optimized for swift responses and common failures, while offering developers and researchers a seamless deployment experience on local setups, providing a foundation for crafting innovative language agents and facilitating real-world evaluations. We elucidate both the challenges and promising opportunities, aspiring to set a foundation for future research and development of real-world language agents.
We welcome contributions from everyone. Before you start, please take a moment to read our CONTRIBUTING.md guidelines for issues and PRs. This will help ensure that your contribution process is smooth and consistent with the projectβs standards.
Join our Discord for help if you encounter any issues with our online demo or local deployment. Alternatively, create an issue if you have trouble with features or code.
- [2023, Oct 26] We've reached 3,000 users! π Heartfelt thanks to all our users and contributors! π Please bear with us as we navigate through unexpectedly high traffic on our servers. We appreciate your patience and ready to assist as soon as possible!
- [2023, Oct 18] Try out our Lemur, the SOTA and open-sourced foundation models for language agents, matching ChatGPT on 15 agent tasks!
- [2023, Oct 17] Check out the OpenAgents paper here!
- [2023, Oct 13] We've released OpenAgents platform code for all three agents, server backend and frontend! Feel free to setup your localhost one, and play with OpenAgents!
- [2023, Aug 17] Our platform has officially reached 500 users! π
- [2023, Aug 8] We've released OpenAgents demos, including Data, Plugins, and Web agents! Check tutorials and use cases!
We built three real-world agents with chat-based web UI as demonstration(check OpenAgents demos). Here is a brief overview of our OpenAgents platform. You can find more details about concepts & designs in our documentation.
Data Agent is a comprehensive toolkit designed for efficient data operations. It provides capabilities to:
- π Search: Quickly locate the data you need.
- π οΈ Handle: Streamline data acquisition and processing.
- π Manipulate: Modify data to suit specific requirements.
- π Visualize: Represent data in a clear and insightful manner.
With its proficiency in writing and executing code, Data Agent simplifies a wide range of data-centric tasks. Discover its potential through various use cases.
Plugins Agent seamlessly integrates with over 200 third-party plugins, each handpicked to enrich various facets of your daily life. With these plugins at its disposal, the agent empowers you to tackle a wide range of tasks and activities more efficiently.
π Sample Plugins Include:
- ποΈ Shopping: Klarna Shopping
- βοΈ Weather: XWeather
- π¬ Scientific Exploration: Wolfram Alpha
Harness the power of synergy! Plugins Agent supports the concurrent use of multiple plugins. Planning a trip? Seamlessly integrate functionalities from Klook, Currency converter, and WeatherViz.
Simplify your choices with our Auto Plugin Selection feature. Let the agent intuitively search and suggest the best plugins tailored to your needs.
Dive into more use cases to see Plugins Agent in action.
Web Agent harnesses the power of a Chrome extension to navigate and explore websites automatically. This agent streamlines the web browsing experience, making it easier to find relevant information, access desired resources, and so on.
Examples of What Web Agent Can Do:
- π Google Maps Navigation: Planning a journey? Simply relay your starting point and destination to Web Agent. It will navigate Google Maps for you and present the best routes.
- π¦ Twitter Postings: Engage in a conversation with Web Agent and wish to share something on Twitter? Mention the content, and Web Agent will handle your tweet effortlessly.
- π Google Form Assistance: Need to sign up for an event or activity? Share the Google Form link and the required details. Web Agent will populate the form for you.
Witness the full potential of Web Agent in these use cases.
We've released the OpenAgents platform code. Feel free to deploy on your own localhost!
Here is a brief system design of OpenAgents:
Please check the following folders and README files to set up & localhost:
- Backend: the flask backend to host our three agents.
- Frontend: the frontend UI and WeBot Chrome extension.
p.s.: We have renamed some arguments in code for better readability. If you have pulled the code before 10/26/2023, just a reminder that if you want to you pull the latest code, previous local chat history will be lost because of different key names.
Before we dive into how to extend OpenAgents, let's first take a glance at the code structure for better understanding. The code structure of OpenAgents is shown below:
βββ backend # backend code
βΒ Β βββ README.md # backend README for setup
βΒ Β βββ api # RESTful APIs, to be called by the frontend
βΒ Β βββ app.py # main flask app
βΒ Β βββ display_streaming.py # rendering the streaming response
βΒ Β βββ kernel_publisher.py # queue for code execution
βΒ Β βββ main.py # main entry for the backend
βΒ Β βββ memory.py # memory(storage) for the backend
βΒ Β βββ schemas.py # constant definitions
βΒ Β βββ setup_script.sh # one-click setup script for the backend
βΒ Β βββ static # static files, e.g., cache and figs
βΒ Β βββ utils # utilities
βββ frontend # frontend code
βΒ Β βββ README.md # frontend README for setup
βΒ Β βββ components # React components
βΒ Β βββ hooks # custom React hooks
βΒ Β βββ icons # icon assets
βΒ Β βββ next-env.d.ts # TypeScript declarations for Next.js environment variables
βΒ Β βββ next-i18next.config.js # configuration settings for internationalization
βΒ Β βββ next.config.js # configuration settings for Next.js
βΒ Β βββ package-lock.json # generated by npm that describes the exact dependency tree
βΒ Β βββ package.json # manifest file that describes the dependencies
βΒ Β βββ pages # Next.js pages
βΒ Β βββ postcss.config.js # configuration settings for PostCSS
βΒ Β βββ prettier.config.js # configuration settings for Prettier
βΒ Β βββ public # static assets
βΒ Β βββ styles # global styles
βΒ Β βββ tailwind.config.js # configuration settings for Tailwind CSS
βΒ Β βββ tsconfig.json # configuration settings for TypeScript
βΒ Β βββ types # type declarations
βΒ Β βββ utils # utilities or helper functions
βΒ Β βββ vitest.config.ts # configuration settings for ViTest
βΒ Β βββ webot_extension.zip # Chrome extension for Web Agent
βββ real_agents # language agents
βββ adapters # shared components for the three agents to adapt to the backend
βββ data_agent # data agent implementation
βββ plugins_agent # plugins agent implementation
βββ web_agent # web agent implementation
As shown, backend/
and frontend/
are self-contained and directly deployable (see here).
It does not mean they cannot be modified.
Instead, you can just follow the conventional client-server architecture to extend the backend and frontend as you wish.
For real_agents/
, we design it to be "one agent, one folder", so that it is easy to extend a new agent.
It is worth noting that we name it "real agents" because not only the conceptual language agent part is included, but also the gaps between the language agent and the backend are filled here.
For example, adapters/
contains the shared adapter components like stream parsing, data model, memory, callbacks, etc.
We refer interested readers to our paper for concepts and implementation designs.
And we thank LangChain as we base on their code to build real agents.
If you want to build a new agent beyond the three agents we provide, you can follow the steps below:
- Refer to the
real_agents/
folder to see how previous agents are implemented, and create a new folder for your agent. - Implement the agent logic in the new folder. Use the components under
adapters/
folder when needed. - Add a
chat_<new_agent>.py
file underbackend/api/
folder to define the chat API for the new agent, which will be called by the frontend. - Register new constants in
backend/schemas.py
if needed. - Add a new
OpenAgentID
infrontend/types/agent.ts
and the corresponding API infrontend/utils/app/api.ts
andfrontend/utils/app/const.ts
. - Implement the agent UI in
frontend/components/Chat/Chat.tsx
andfrontend/components/Chat/ChatMessage.tsx
when needed. - Run localhost script and test your new agent.
Note, if new data types, i.e., beyond text, image, table, and json, you may need to implement its parsing logic in backend/display_streaming.py
and add new data models.
Extending a new LLM as the agent backbone is simpler if the LLM is already hosted and can be called via API.
Just register your new model in backend/api/language_model.py
. Just refer to lemur-chat as a template.
If the LLM is not hosted yet, we have a tutorial on how to deploy a new LLM and expose it as an API here (LLM hosting to todo).
If you want to extend a new tool in Plugins Agent, you can follow the steps below:
- Refer to the already built plugins in
real_agents/plugins_agent/plugins/
, and create a new folder for your tool. - Implement the tool logic in the new folder. Note that
ai-plugin.json
andopenapi.yaml
are essential for the tool to be recognized(which can be generated by LLM following the others rather than manually written). And thepaths/
are for the actual tool API call. - Register the new tool name in
real_agents/plugins_agent/plugin_names.py
.
Thanks to open-sourced communitiesβ efforts, such as LangChain, ChatBot UI, Taxy.ai browser extension and others. We are able to build our interface prototype much more conveniently and efficiently.
We welcome contributions and suggestions, together we move further to make it better! Following the steps will be well-received:
- Step1: Post an issue if you want to add any additional features, enhancements, or encounter any problems during your experience. We would appreciate it if you follow the issue template. The issues will be discussed and assigned there.
- Step2: Whenever an issue is assigned, you can contribute by creating a Pull Request by following the PR template. You can also claim for any open issues. Together we can make OpenAgents better!
- Step3: PR will be merged or iterated after review and discussion. Thanks for your contribution!
Before you start, we highly recommend taking a moment to check here before contribution.
Please check here for full documentation, which will be updated to stay on pace with the demo changes and the code release.
Heartfelt appreciation to Ziyi Huang, Roxy Rong, Haotian Li, Xingbo Wang, Jansen Wong, and Chen Henry Wu for their valuable contributions to the OpenAgents. Their expertise and insights were instrumental in bringing this project to fruition!
Thanks to all the contributors!
If you find our work helpful, please cite us:
@misc{OpenAgents,
title={OpenAgents: An Open Platform for Language Agents in the Wild},
author={Tianbao Xie and Fan Zhou and Zhoujun Cheng and Peng Shi and Luoxuan Weng and Yitao Liu and Toh Jing Hua and Junning Zhao and Qian Liu and Che Liu and Leo Z. Liu and Yiheng Xu and Hongjin Su and Dongchan Shin and Caiming Xiong and Tao Yu},
year={2023},
eprint={2310.10634},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
We would like to thank Google Research, Amazon AWS, and Salesforce Research for their research gift funds to this open-source effort!