sensei

Sensei

The documentation agent for coding agents.

Sensei searches multiple authoritative sources, cross-validates, and synthesizes accurate answers so your AI writes working code on the first try.

Try it live →

Install

Claude Code

claude plugin marketplace add 803/sensei
claude plugin install --scope user sensei@sensei-marketplace

Other MCP Clients

Remote (recommended):

https://api.sensei.eightzerothree.co/mcp

Local:

uvx sensei-ai --help

API

curl -X POST https://api.sensei.eightzerothree.co/query \
  -H "Content-Type: application/json" \
  -d '{
    "query": "How do I authenticate with OAuth?",
    "language": "python",
    "library": "fastapi"
  }'

Why Sensei

20x more context efficiency

Other tools paste raw docs into your context window—100,000 to 300,000 tokens of unfiltered content. Sensei reads, validates, and synthesizes. You get 2,000-10,000 focused tokens. Your agent's context stays clean for the actual work.

Optimized research methodology

Sensei researches like a senior engineer. It goes wide first to survey options, then deep on promising paths. It follows a trust hierarchy—official docs → source code → real implementations → community content—and matches sources to goals. Complex questions get decomposed into parts, researched separately, and synthesized into one answer you can trust.

Continuous improvement

Your agent gives feedback to Sensei. Did the code work? Was the guidance correct? Every outcome is a verified reward signal. We fine-tune the model from real results. Success reinforces what works. Failure refines what doesn't.

The Tools

Alongside third-party tools like Context7 and Tavily, Sensei includes three purpose-built tools:

Kura — Knowledge cache. First query: thorough research across all sources. Every query after: instant. Complex questions get decomposed into parts—and each part gets cached as a reusable building block. Future questions that share parts get faster, more accurate answers.

Scout — Source code exploration. Glob, grep, and tree any public repository at any tag, branch, or commit SHA. Local clones created on-demand. When docs are unclear, read what the code actually does.

Tome — llms.txt ingestion. llms.txt is the future of AI-readable documentation. Tome ingests on-demand from any domain and saves for future use. Official docs, formatted for agents, always available.

For Teams

Bring your own sources. Internal wikis. Private repos. Proprietary APIs. Connect them via MCP, and Sensei searches them alongside everything else.

Self-host the full stack. Sensei runs on your infrastructure. Your queries stay on your network. Complete control when you need it.

Open source. Inspect it. Fork it. Trust it.

Contributing

uv sync --group dev
uv run pre-commit install

See CONTRIBUTING.md for development guidelines.

License

MIT

---

Built with PydanticAI and FastMCP