openagent

An end-to-end AI agent platform for building, orchestrating, publishing, and operating AI applications.
Flask + LangChain/LangGraph backend, Vue 3 workspace, visual workflows, datasets, tools, and OpenAPI delivery.

Sponsored by Atlas Cloud and Bloome

Visit Website · API Docs · 中文文档 · GitHub

Table of Contents

• About The Project
• Architecture
• Built With
• Getting Started
• Usage
• Testing
• Contact
• Sponsored By
• Acknowledgments

About The Project

OpenAgent is a full-stack platform for teams building AI applications rather than a single chat demo. The repository combines a Flask backend, Celery workers, a Vue 3 frontend, visual workflow authoring, dataset and document management, public app and workflow publishing, and OpenAPI-based delivery.

What the current codebase already supports:

• Use the home assistant to route user requests to published public agents through A2A, or turn natural-language requirements into new AI app creation flows.
• Build and manage AI apps from a dedicated workspace with draft, publish, analysis, version comparison, and prompt comparison flows.
• Enable Deep Research to let the app decompose complex tasks and coordinate bound capabilities across multi-step execution, suitable for scenarios requiring deep reasoning and concrete outputs.
• Design workflows visually with nodes for LLMs, tool calls, dataset retrieval, code execution, HTTP requests, branching, text processing, template transforms, and structured parameter extraction.
• Manage datasets, upload documents, inspect segments, and connect retrieval to agents and workflows.
• Browse public apps, tools, and workflows through store-style views.
• Expose published apps over REST and SSE through POST /api/openapi/chat.

Architecture

Click the diagram to view the full-resolution architecture image.

See the Documentation index for source-level guides and the Chinese documentation index.

Detailed source-level guide: Architecture documentation.

Built With

• AI framework and orchestration: LangChain, LangGraph, workflow orchestration, tool calling, A2A delegation, skills, memory
• Knowledge and retrieval: RAG, semantic retrieval, full-text retrieval, hybrid retrieval, Weaviate, FAISS
• Backend: Python, Flask, SQLAlchemy, Celery, Flask-SocketIO, Redis, PostgreSQL
• Frontend: Vue 3, JavaScript / TypeScript, Vite, TailwindCSS, Pinia, Vue Flow, Arco Design
• Infrastructure and delivery: Docker Compose, Nginx, OpenAPI, SSE
• Model integrations: OpenAI, Atlas Cloud, DeepSeek, Grok, Google, Moonshot, Tongyi, Wenxin, Ollama, Zhipu

Sponsored By

Atlas Cloud

• Atlas Cloud is a model provider on OpenRouter. Integrate it in this project via ATLASCLOUD_API_KEY and ATLASCLOUD_API_BASE.
• Official website: Atlas Cloud
• Integration docs: https://www.atlascloud.ai/docs

Bloome

• Bloome is an AI-agent IM platform where you can interact with AI agents and build collaborative communities. Official website: Bloome

(back to top)

Getting Started

Prerequisites

• Docker 20.10+
• Docker Compose 2.x
• 8 GB+ RAM recommended for the full stack
• Access to at least one supported model provider API key

Installation

1. Clone the repository.

   git clone https://github.com/Haohao-end/openagent.git
   cd openagent
   

2. Create the runtime environment file.

   cp api/.env.example api/.env
   

3. Review the minimum required settings in api/.env.

   • JWT_SECRET_KEY
   • POSTGRES_PASSWORD
   • REDIS_PASSWORD
   • WEAVIATE_API_KEY
   • VITE_API_PREFIX
   • At least one provider key such as OPENAI_API_KEY, ATLASCLOUD_API_KEY, DEEPSEEK_API_KEY, or DASHSCOPE_API_KEY

4. Start the Docker stack.

   cd docker
   docker compose up -d --build
   

5. Open the local services.

   Service	URL	Notes
   Frontend	http://localhost:3000	Vue 3 web UI
   API	http://localhost:5001	Flask REST API
   Nginx	http://localhost	Reverse proxy

Local Development

For local development, start only the supporting Docker services; the API and UI run directly on your machine. This avoids building or starting the API, UI, Celery, or Nginx containers:

cd docker
docker compose up -d --no-build llmops-db llmops-redis llmops-weaviate
docker compose ps llmops-db llmops-redis llmops-weaviate

Use localhost and the published ports for PostgreSQL, Redis, and Weaviate in your local api/.env configuration. See Docker setup for details. Keep full-stack docker compose up -d --build for containerized deployment, not this local workflow.

Backend:

cd api
pip install -r requirements.txt
flask run --port 5001

Frontend:

cd ui
npm install
npm run serve

Vite serves the frontend on port 5173 by default. The frontend configuration resolves the API base from VITE_API_PREFIX, and local development commonly proxies /api to the Flask backend.

Useful commands:

cd api
pytest

cd ui
npm run type-check
npm run lint
npm run build
npm run test:unit -- --run

(back to top)

Usage

1. Home Assistant Experience

Use the home page as the default assistant entry point to route user questions to the most relevant published public agents through A2A, or describe a new idea in natural language and trigger AI app creation. The same surface also supports multi-turn chat, suggested prompts, image upload, and audio input.

2. App Workspace and Deep Research

The app workspace is the main work area for an AI app, not a standalone settings page. The left side handles model, prompt, and capability bindings. The right side is used for live debugging, execution traces, and result checks. In the current codebase, the README term Deep Research maps to the deep thinking mode behind enable_deep_thinking.

Key capabilities:

• Configuration and version management: manage model changes, prompt logic, drafts, publishing, version comparison, prompt comparison, and app duplication in one place.
• Capability integration: bind plugins, MCP, Skills, child Agent apps, workflows, and datasets through a single workspace.
• Complex task execution: with Deep Research enabled, the app can break work into steps and coordinate bound capabilities across a longer execution chain.
• Sandbox and artifact output: support script execution, code handling, file generation, and attachment export for tasks that need concrete outputs.
• Debugging and result verification: use the right-side panel to run real conversations and inspect deep execution timelines, task state, produced artifacts, and final results.

3. Visual Workflow Editor

Author workflows with nodes such as LLM, tool, dataset retrieval, code, HTTP request, template transform, text processor, variable assigner, parameter extractor, if/else, start, and end.

4. Dataset and Retrieval

Create datasets, upload documents, inspect segments, and wire retrieval nodes into workflows or AI apps for knowledge-enabled behavior.

5. OpenAPI Delivery

Publish an app and call it over POST /api/openapi/chat with standard or streaming responses, including support for multi-turn conversation identifiers.

(back to top)

Testing

The repository already includes automated backend and frontend tests.

• Backend: cd api && pytest
• Frontend unit tests: cd ui && npm run test:unit -- --run
• Frontend type check: cd ui && npm run type-check
• Frontend build validation: cd ui && npm run build

Contact

• Project Link: https://github.com/Haohao-end/openagent
• Website: https://openllm.cloud
• API Docs: https://s.apifox.cn/c76bd530-fd50-429c-94cc-f0e41c2675d1/api-305434417
• DeepWiki: https://deepwiki.com/Haohao-end/openagent

Acknowledgments

• Thanks to Atlas Cloud and Bloome for supporting OpenAgent.
• Special thanks to Rui Yang and Haoyu Wang (Johns Hopkins University) for responsibly reporting a Host Header poisoning issue in the built-in tool icon URL construction and helping improve the security of this project.
• Special thanks to Ryan B for responsibly reporting an SSRF vulnerability in the custom API-tool server URL and helping improve the project’s outbound request security.
• Thanks to 汪钶楠 (Wang Kenan) for reporting an AppleScript argument-handling concern in the macOS screenshot helper and helping us verify a safe fix.

(back to top)