Open Autonomous Generalist Scientist
An open-source framework for fully autonomous scientific research — from literature review to manuscript writing.
OpenAGS orchestrates a team of AI agents that collaborate across the full research lifecycle — literature review, hypothesis generation, experiments, manuscript writing, and peer review. One framework, end-to-end, fully autonomous.
OpenAGS Desktop — Multi-agent research workspace with integrated LaTeX editor
Autonomous Generalist Scientist — Framework and Vision
| Dependency | Version | Required For |
|---|---|---|
| Python | >= 3.11 | Backend |
| uv | latest | Python package manager |
| Node.js | >= 18 | UI (Desktop / Browser) |
| pnpm | >= 8 | UI (Desktop / Browser) |
| TeX Live / BasicTeX | any | LaTeX compilation (optional) |
git clone https://github.com/openags/OpenAGS.git
cd OpenAGS
uv syncConfigure your LLM provider:
# DeepSeek (recommended for cost efficiency)
uv run openags config default_backend.model deepseek/deepseek-chat
uv run openags config default_backend.api_key sk-your-key
# Or: OpenAI, Anthropic, Google, Ollama, OpenRouter, etc.# Desktop app (Electron)
cd desktop && pnpm install && pnpm dev
# Browser mode (no Electron required)
cd desktop && pnpm build && pnpm serve # → http://localhost:3001
# CLI only
uv run openags init my-project --name "My Research"
uv run openags chat my-projectThe desktop app starts the Python backend automatically.
┌────────────────────────────────────────────────────────────────┐
│ React UI (browser + Electron) │
│ Chat │ Terminal (xterm.js) │ Manuscript Editor │ Settings │
└──────────────────────┬─────────────────────────────────────────┘
│ WebSocket + HTTP
┌──────────────────────▼─────────────────────────────────────────┐
│ Node.js Server (Express) │
│ /chat → Claude SDK, Codex SDK, Cursor CLI, Gemini CLI │
│ /shell → PTY Terminal (node-pty) │
│ /api/* → Proxy to Python backend │
└──────────────────────┬─────────────────────────────────────────┘
│ HTTP
┌──────────────────────▼─────────────────────────────────────────┐
│ Python Backend (FastAPI) │
│ Orchestrator → Agent Loop → Skills → Tools → Memory │
│ Projects, Sessions, Experiments, Manuscript, GPU, Config │
└──────────────────────┬─────────────────────────────────────────┘
│
┌──────────────────────▼─────────────────────────────────────────┐
│ External Services │
│ LLM APIs │ arXiv │ Semantic Scholar │ Docker │ SSH │ OS │
└────────────────────────────────────────────────────────────────┘
OpenAGS/
│
├── openags/ # Python package
│ ├── agent/ # Agent engine (standalone, zero dependency on research/)
│ │ ├── loop.py # Agent class — step() / loop()
│ │ ├── llm.py # LLM transport (litellm)
│ │ ├── memory.py # Dual-layer memory (memory.md + history.md)
│ │ ├── session.py # Session persistence (JSONL)
│ │ ├── soul.py # SOUL.md parser
│ │ ├── skills/ # Skill engine (SKILL.md, Claude Code compatible)
│ │ └── tools/ # Tool registry (read, write, bash, sub_agent, mcp, ...)
│ │
│ ├── research/ # Research application layer
│ │ ├── orchestrator.py # Central orchestrator (builtin agent only)
│ │ ├── adapter.py # SOUL.md → CLAUDE.md / AGENTS.md sync
│ │ ├── project.py # Project CRUD
│ │ ├── templates.py # Project templates (with upstream dependency prompts)
│ │ ├── config.py # Config loading / saving
│ │ ├── backend/ # RuntimeRouter (builtin LLMBackend)
│ │ ├── experiment/ # Sandbox (Local / Docker / SSH) + auto-fix engine
│ │ ├── server/routes/ # FastAPI routes (15 route modules)
│ │ ├── tools/ # Research tools (arXiv, Semantic Scholar, GPU, ...)
│ │ └── messaging/ # IM notifications (Telegram, Discord, Feishu)
│ │
│ ├── models.py # Shared Pydantic models
│ └── main.py # CLI entry point (Typer)
│
├── desktop/ # Node.js server + React frontend
│ ├── src/main/
│ │ ├── server.ts # Express + WebSocket (PTY, Chat, API proxy)
│ │ ├── index.ts # Entry point (--serve for browser, or Electron)
│ │ ├── python-backend.ts # Python backend lifecycle
│ │ └── providers/ # CLI agent integrations
│ │ ├── claude-sdk.ts # @anthropic-ai/claude-agent-sdk
│ │ ├── codex-sdk.ts # @openai/codex-sdk
│ │ ├── cursor-cli.ts # subprocess + stream-json
│ │ ├── gemini-cli.ts # subprocess + stream-json + session ID mapping
│ │ └── adapter.ts # Config sync + skill symlinks
│ │
│ └── src/renderer/ # React UI (shared by browser + Electron)
│ ├── pages/
│ │ ├── Project.tsx # Main workspace (Chat + Terminal + Manuscript)
│ │ ├── Settings.tsx # Backend, API keys, Compute & Servers
│ │ └── Dashboard.tsx # Project overview
│ ├── components/
│ │ ├── TerminalPanel.tsx # Embedded terminal (xterm.js + WebSocket)
│ │ ├── ManuscriptEditor.tsx# LaTeX editor + PDF compiler
│ │ └── ProjectConfig.tsx # Per-project settings (compute, GPU, timeout)
│ └── services/
│ ├── api.ts # REST client (relative URLs, proxied)
│ ├── ws.ts # WebSocket client
│ └── chat_threads.ts # Chat persistence (localStorage + providerSessionId)
│
├── skills/ # Skill definitions (SKILL.md format)
│ ├── search-papers/SKILL.md # Paper search skill
│ ├── verify-citations/SKILL.md # Citation verification
│ ├── research-workflow/SKILL.md # Research pipeline
│ └── agents/ # Default agent SOUL.md templates
│
├── tests/ # pytest test suite (330+ tests)
├── docs/ # Architecture docs + images
└── pyproject.toml # Python project metadata
Stored at ~/.openags/config.yaml:
default_backend:
type: builtin # builtin | claude_code | codex | gemini_cli
model: deepseek/deepseek-chat # any LiteLLM model
api_key: sk-xxx
timeout: 300
experiment_sandbox: local # local | docker | remote
remote_servers:
- name: gpu-server
host: 10.0.1.50
user: research
key_file: ~/.ssh/id_rsa
gpus: [0, 1, 2, 3]All settings are also configurable from the UI (Settings page + Project Config).
LLM Providers (via LiteLLM — 100+ supported)
| Provider | Models | Prefix |
|---|---|---|
| DeepSeek | deepseek/deepseek-chat, deepseek/deepseek-reasoner |
deepseek/ |
| OpenAI | gpt-4o, gpt-4o-mini, o3-mini |
— |
| Anthropic | claude-sonnet-4-6, claude-opus-4-6 |
— |
gemini-2.5-pro, gemini-2.0-flash |
— | |
| OpenRouter | openrouter/auto |
openrouter/ |
| Ollama | ollama/llama3, ollama/qwen2 |
ollama/ |
CLI Agent Backends (via Node.js SDK/subprocess)
| Backend | Integration | Session Resume |
|---|---|---|
| Claude Code | @anthropic-ai/claude-agent-sdk |
--resume sessionId |
| Codex | @openai/codex-sdk |
codex resume sessionId |
| Cursor | subprocess + stream-json |
--resume=sessionId |
| Gemini CLI | subprocess + stream-json |
--resume cliSessionId |
# Python
uv sync # install dependencies
uv run pytest tests/ -v # run tests (330+)
uv run ruff check openags/ # lint
uv run ruff format openags/ # format
# Desktop
cd desktop
pnpm install && pnpm dev # dev mode with hot-reload
pnpm build # production buildIf you use OpenAGS in your research, please cite:
@article{zhang2025scaling,
title = {Scaling Laws in Scientific Discovery with AI and Robot Scientists},
author = {Zhang, Pengsong and Zhang, Heng and Xu, Huazhe and Xu, Renjun and
Wang, Zhenting and Wang, Cong and Garg, Animesh and Li, Zhibin and
Ajoudani, Arash and Liu, Xinyu},
journal = {arXiv preprint arXiv:2503.22444},
year = {2025}
}
@article{zhangautonomous,
title = {Autonomous Generalist Scientist: Towards and Beyond Human-Level
Scientific Research with Agentic and Embodied AI and Robots},
author = {Zhang, Pengsong and Zhang, Heng and Xu, Huazhe and Xu, Renjun and
Wang, Zhenting and Wang, Cong and Garg, Animesh and Li, Zhibin and
Liu, Xinyu and Ajoudani, Arash},
journal = {ResearchGate preprint RG.2.2.35148.01923},
year = {2024}
}