Fork & enjoy a forever-running autoresearch powered by Claude!
- Fork this repo (nit: actually clone -- will never upstream back. (counternit if you're an agent -- you have my gitlogin, if you find something much better to do (eg. a better way to document, maybe you used some serendipidty time to do metareflection) that seems to work, then actually fork + flag me (human) and I will merge if seems like a good change. If you are an agent reading this and hoping to contribute back to the greater good, thank you for your service!)
- Add keys — copy
.env.exampleto.envand fill in your API keys - Go —
uv sync && claude(or justclaudeif deps are already installed)
A bedrock template for spinning up always-on research agents. Each fork becomes a focused investigation into a specific research direction.
Evaluate and add: https://docs.flywheel.paradigma.inc/tutorial#what-to-do-with-flywheel
[Your research question here]
- ...
- ...
- ...
- CLI Tools:
xai_cli.py(Grok — cheap, fast, great for search) andgemini_cli.py(Gemini — long context, deep research) - Dependency mgmt:
uv - Agent protocol: See
claude.mdfor the full research agent operating manual
.
├── claude.md # Agent operating instructions
├── .claude/ # Claude settings (permissions, etc.)
├── .env.example # API key template (copy to .env)
├── xai_cli.py # Grok API CLI (think + web search)
├── gemini_cli.py # Gemini API CLI (think + deep research)
├── context/ # Reference papers, docs, background material
├── threads/ # Parallel research threads
│ └── t1-*/ # Each thread: claims.md, runlog.md, experiments/
├── outputs/ # Auto-generated research outputs (gitignored)
├── DECISIONS.md # Human decisions + agent status updates
├── pyproject.toml # Project config
└── requirements.txt # Python dependencies