Librarian MCP

pip install librarian-mcp

A real, measured alternative to "bigger context windows." Pre-curated canonical memory + prose/code provenance checking + benchmark metrics, delivered as a Model Context Protocol server that works across Claude Code, Cursor, VSCode (via Continue), and any MCP-capable client.

Try it without installing →

What it does

Five tools, all exposed via MCP:

Why we built this

Independently measured result (Eyewitness Benchmark R10, April 2026, eight models across four vendors, 1,200 graded calls, inter-rater kappa 0.883/0.850):

• Without the Librarian (COLD): mean 8.7% correct

• With the Librarian (HOT): mean 94.8% correct — 86.1 percentage-point lift

• Haiku 4.5 (cheapest) ties Opus 4.7 (most expensive) at 19x cost difference

• 4.3x more right answers per dollar of compute

Applied inside Microsoft Copilot's inference path, the same architecture recovers an estimated $750M/year in waste. Inside Anthropic's developer tools, ~$130M/year. Full methodology in the R9 Empirical Test Companion Paper.

librarian_context — Intent API

librarian_context(intent="outreach", max_tokens=16000)

List inputs for union queries: intent='["benchmark", "founder_voice"]'

Returns:

{
  "packet": "...markdown...",
  "sections_included": ["r9v2_base.md", "canonical/canonical_values.yaml", ...],
  "token_count": 14832,
  "source_version": "a1b2c3d4e5f6",
  "truncation_note": null
}

metrics_summary — Schema

{
  "total_calls": 1200,
  "per_vendor": {
    "anthropic": {
      "calls": 600,
      "hot_accuracy": 95.3,
      "cold_baseline_est": 8.2,
      "dollars_saved_est": 42.17,
      "cache_hit_rate": 50.0
    }
  },
  "per_model": {
    "claude-haiku-4-5-20251001": { "..." : "..." }
  },
  "cumulative_hot_accuracy": 94.8,
  "cumulative_cold_baseline_est": 8.7,
  "cumulative_dollars_saved_est": 127.50,
  "opt_in_share": false,
  "since": "all_time"
}

Pricing

The commercial tiers pay for the commons. No grant funding, no VC, no extractive margin. Cost+20% on operating expense. That's it.

Why MCP (not a Cursor extension)

Because you shouldn't have to pick between your AI assistants. MCP servers work across Claude Code, Cursor (v0.45+), Continue (VSCode / JetBrains), Zed, and every MCP-capable client in the roadmap. One server, all your tools.

Install

Quick start (local, Python 3.10+)

git clone https://github.com/liana-banyan/librarian-mcp.git
cd librarian-mcp
pip install -e .
librarian-mcp  # starts on stdio for MCP clients

With optional dependencies

pip install -e ".[all]"   # tiktoken (accurate token counts) + anthropic + pyyaml
pip install -e ".[dev]"   # + pytest, ruff, mypy for development

Claude Code

claude mcp add librarian python -m librarian_mcp

Cursor

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "librarian": {
      "command": "python",
      "args": ["-m", "librarian_mcp"]
    }
  }
}

Continue (VSCode / JetBrains)

See docs/continue-integration.md.

Development

pip install -e ".[dev,all]"
ruff check src/ tests/          # lint
mypy --strict src/librarian_mcp/  # type check
pytest -v                        # test (34 tests)

Status

April 21, 2026 — v0.2.0. Intent-aware librarian_context live with bundled preload (R10-validated). Benchmark metrics recording live. Prose Provenance tool upgraded to v0.2.0. PyPI name librarian-mcp reserved. CI/CD staged.

License

AGPL-3.0. Commercial licensing for the paid tiers is a separate agreement; the Pledged Commons tier is covered by AGPL + the Cooperative Defensive Patent Pledge.

Contact

• General: hello@liana-banyan.com

• Enterprise: enterprise@liana-banyan.com

• Press / AI policy / datacenter-alternative questions: press@liana-banyan.com

• Founder: Jonathan Jones, Founder & General Manager, Liana Banyan Corporation (Wyoming C-Corp)

Contributing

We welcome contributions — code, corpus preloads, benchmark replications, and research extensions.

• BOUNTIES.md — paid bounties for specific contributions, from $25 good-first-bounty issues to $500 deep bounties

• BUILDING_TOGETHER.md — guide to running, extending, and contributing back upstream

"You build the Features — We're building the Board."

Pledged into the commons. For the Keep.