legal-mcp

A remotely callable MCP server for US legal research. Baseline ("套壳") milestone: a thin, source-agnostic wrapper over CourtListener (the Free Law Project's 9M+ federal and state court opinions — free and public), exposed over the MCP streamable-HTTP transport.

It is deliberately the skeleton the value layers plug into later (multi-jurisdiction aggregation, full-text retrieval, citation graphs, freshness monitoring), mirroring the architecture of paper-mcp.

Design discipline

The server returns only primary-source text and citation metadata retrieved from the upstream API. It never lets a model generate legal content — agents build evidence packs they can check and footnote. (Legal hallucination has real sanctions; the server refuses to be the source of one.)

Tools

Architecture

legal_mcp/
├── __init__.py            version string
├── models.py              LegalDoc — normalized, source-agnostic record
├── aggregate.py           pure RRF fusion + de-dup (by citation / case name)
├── server.py              FastMCP bootstrap + tool registrations
└── sources/
    ├── base.py            LegalSource Protocol (the swappable contract)
    ├── __init__.py        source registry + aliases
    └── courtlistener.py   CourtListener REST client → LegalDoc

Adding a jurisdiction (CAP, GovInfo, Congress, eCFR, OpenStates) is a new sources/<name>.py implementing the LegalSource contract plus one line in the registry — no server change.

Configuration

Run

pip install -e .
legal-mcp                 # serves streamable-HTTP at 127.0.0.1:9500/mcp

Anonymous access already allows search and document fetch (rate-limited). Set COURTLISTENER_API_TOKEN for higher limits and citation lookup.

Roadmap

1. MVP (this) — CourtListener case law: search, full text, citations.

2. Add federal sources — eCFR (regulations), GovInfo (US Code), Congress (bills/votes/members), Federal Register, 26 USC/CFR (tax).

3. Add CAP (Harvard) historical 50-state case law, fused with CourtListener.

4. Add OpenStates (50-state legislation) for the state layer UK has no analogue for.

5. Freshness pipeline — incremental watermark crawl + embeddings (slow data: ~hundreds of new docs/day, <10 min GPU/day).