Legal Search MCP

Legal Search MCP

English | 한국어

Korean case law, statutes and sentencing data as MCP tools. Connect an AI client to find judgments, look up statutes as of a date, check what a judgment held in its own words and calculate statutory and guideline sentencing ranges.

Four read-only tools:

This repository publishes the MCP tools used by 로풀 (Lawful), a free Korean legal-AI service. The web application and hosted Agent API are maintained separately. The research prototype and benchmark archive live in legal_mcp — see Paper.

Quick start — hosted

The hosted service provides the full corpus, including statute amendment history, with no rate limits. Three ways to connect:

Claude / ChatGPT on the web — no code. Add a custom connector with this URL, sign in with OAuth when asked, and the four tools appear. No API key needed:

https://mcp.crow-tit.com/mcp

Step-by-step with the exact menu clicks: crow-tit.com/docs#mcp

Claude Code — one command, with a free key from console.crow-tit.com:

claude mcp add --transport http lawful https://mcp.crow-tit.com/mcp \
  --header "Authorization: Bearer ct_..."

Clients with HTTP MCP configuration — use the same key. Configuration formats vary; for clients that accept an mcpServers HTTP entry:

{
  "mcpServers": {
    "lawful": {
      "url": "https://mcp.crow-tit.com/mcp",
      "headers": { "Authorization": "Bearer ct_..." }
    }
  }
}

Quick start — self-host

Requires Python 3.11+. Install in a virtual environment; the bundled sample works without API keys:

git clone https://github.com/LimEulYoung/lawful-mcp
cd lawful-mcp
python3 -m venv .venv
.venv/bin/python -m pip install -e .

For a local stdio client, replace the path below with the absolute path to your clone. The client starts the server:

{
  "mcpServers": {
    "lawful-mcp": {
      "command": "/absolute/path/to/lawful-mcp/.venv/bin/lawful-mcp"
    }
  }
}

For HTTP at http://127.0.0.1:8100/mcp, run:

.venv/bin/lawful-mcp --transport http --port 8100

The self-hosted server has no authentication or usage metering. Access control for a shared HTTP endpoint belongs to your deployment.

Configuration and data flow

Settings are listed in .env.example. The server reads process environment variables; it does not load .env automatically. Export values in the launching shell or set them in your MCP client's env configuration.

With the defaults, three tools search and calculate locally. precedent_dive additionally sends the selected public judgment text, case metadata and the supplied question to your configured model endpoint. Keep that question limited to issues in the public judgment.

export DIVE_API_KEY=...
export DIVE_BASE_URL=https://api.openai.com/v1   # any OpenAI-compatible endpoint
export DIVE_MODEL=...

All three values are required to register precedent_dive. Without them, the other three tools still work. Restart the server after changing settings.

precedent_search also takes an optional issue: the legal question the search is meant to answer. With JEV_API_KEY (an OpenRouter key) set, it sends each candidate's result block — public case metadata, holding and preview — and that question to OpenRouter's Jev endpoint, and returns the cases that decided the question first. Without the key, or when a request fails or takes longer than 1.5 seconds, results keep their keyword order.

Optional dense search (USE_DENSE=1) sends search queries to an embedding endpoint and, when reranking is enabled, queries and candidate excerpts to a reranking endpoint. It requires the dense extra and corpus vectors, which the bundled sample does not contain.

Check the installation

The tests use the bundled sample and constructed test data; no network or model keys are needed:

.venv/bin/python -m pip install -e ".[dev]"
.venv/bin/python -m pytest tests/ -q

The corpus

The tools read one SQLite file. The bundled sample is about 48 MiB:

The sample supports local trials and testing. Use the hosted service for the full corpus, or set CORPUS_DB to a compatible database you maintain.

Statute lookups distinguish repeal from loss of effect, such as expiry, when the corpus supplies that information. The sample's lapse register is currently empty; expiry and reinstatement are covered by separate test data.

Default search combines trigram and Kiwi morpheme indexes using reciprocal rank fusion. The morpheme index makes two-character charge names such as 사기, 절도 and 폭행 searchable.

Judgments are non-copyrightable under Article 7 of the Korean Copyright Act. Court and case-number provenance is preserved in the data.

Building your own

scripts/build_sample_db.py extracts a sample from an existing compatible corpus and rebuilds its search indexes. It also documents the schema; it does not download the hosted corpus:

.venv/bin/python scripts/build_sample_db.py --source /path/to/corpus.db \
    --dest my_sample.db --cases 5000 --statutes all

Paper

The retrieval and tool-use design was evaluated on the Korean Bar Examination:

Agentic RAG for Legal Question Answering in Civil Law: Evidence From the Korean Bar Examination Eul Young Lim and Jihun Park IEEE Access, vol. 14, pp. 124441–124458, 2026. doi:10.1109/ACCESS.2026.3722717 — open access

Benchmark code, questions and per-model results are archived in legal_mcp. If you use this tool in research, please cite the paper (CITATION.cff).

Also available

• Lawful Agent API — one question in, a grounded answer out. crow-tit.com
• 로풀 (Lawful) — the free consumer service, 법률AI: chat, search, labour calculators, legal document drafting. lawful.crow-tit.com

Notes

This is a search tool over public legal sources, not legal advice. What it returns is source material and statistics; deciding what they mean for a particular matter is a lawyer's job.

Development happens in a private repository and lands here in batches, so issues are welcome but pull requests may be merged by hand rather than through the button.

Report vulnerabilities privately using SECURITY.md. Code is MIT licensed.