Legal MCP Server

An MCP server that turns Claude into a working legal assistant for Indian law — case-law research with verified citations, limitation and deadline arithmetic, document drafting, matter tracking, and search over your own contracts and case files.

Built on Red Hat's template-mcp-server (FastMCP + FastAPI, OAuth, PostgreSQL, Containerfile, OpenShift, CI).

Defaults to Maharashtra / Bombay High Court; every tool takes an explicit jurisdiction where it matters.

The problem this is built around

An LLM asked a legal question will produce a fluent answer containing citations that look exactly like real ones. Some of them will not exist. In legal work that is not a rough edge — it is the whole risk.

So the design rule throughout is: the server distinguishes what it knows from what it does not, and never closes the gap on its own.

• Case law comes from a real lookup against a real corpus, or the tool reports unavailable. "The source could not be consulted" is never rendered as "no authority exists" - and an empty result over an unsynced year range says so.

• verify_all_citations sweeps any draft or memo and marks every citation VERIFIED, NOT_FOUND, AMBIGUOUS or UNCHECKED. Nothing is quietly dropped or reworded.

• Statutory text is labelled authentic or summary. A curated paraphrase is never presentable as the words of a statute.

• A section missing from a partially bundled Act reports "not in the corpus", not "does not exist".

• Deadline results carry a calendar_confidence block saying which court holidays were actually accounted for.

• Templates refuse to render with a placeholder where a fact should be.

Related MCP server: LegalMCP

What it does

40 tools in seven groups.

Research and citations

search_case_law · get_judgment · search_within_judgment · find_related_proceedings · verify_citation · verify_all_citations · build_research_memo · sync_case_law · case_law_status

Free case law, no API key. Backed by the AWS Open Data release of Indian judgments (CC-BY-4.0): the Supreme Court from 1950 with official S.C.R. and neutral citations, plus 25 High Courts / 45 benches (~17.8M judgments).

sync_case_law downloads the Parquet metadata for the courts and years you care about - tens of MB, not the 1.25 TiB of PDFs - and search then runs locally through DuckDB: instant, offline, unlimited, free. Judgment PDFs are fetched individually from public S3 only when you open one, and cached.

Because the Supreme Court dataset carries real citation fields, verify_citation matches [2024] 10 S.C.R. 108 and 2024INSC735 against the official record rather than guessing from a text search.

Statutes and the new criminal codes

get_section · search_statute · map_criminal_code_section · which_criminal_code_applies · list_bundled_acts

2,040 sections across 19 Acts bundled offline — free, fast, no network. The BNS/BNSS/BSA came into force on 1 July 2024, and the date of the offence, not today's date, decides which code governs. The concordance covers 103 of the most-used provisions in both directions and flags the ones that are substantive changes rather than renumberings (CrPC 200 → BNSS 223, Evidence 65B → BSA 63).

Limitation and deadlines

compute_limitation · find_limitation_rule · list_limitation_rules · compute_cheque_bounce_timeline · compute_deadline · get_court_holidays

Limitation Act 1963 with sections 4, 12(1), 12(2), 14 and 18 modelled — an acknowledgment before expiry restarts the clock, one after it does not revive a time-barred claim, and the tool says which rule it applied and why. compute_cheque_bounce_timeline handles all three section 138 clocks and is loud about the fact that the 30-day notice period cannot be extended.

Matters and hearings

create_matter · update_matter · list_matters · get_matter · add_hearing · list_upcoming_hearings · log_matter_event · get_matter_timeline

Your matters in your own PostgreSQL. list_matters surfaces limitation alerts unprompted.

Your documents

ingest_document · search_my_documents · list_my_documents · get_document · extract_clauses · review_contract

PDF/DOCX/TXT/MD ingest, chunked on the document's own clause boundaries. Hybrid search (Postgres full-text + pgvector, reciprocal-rank fused). review_contract applies India-specific rules — section 27 on post-termination non-competes, section 28 on jurisdiction clauses, seat-versus-venue in arbitration clauses, unstamped instruments — each citing the provision it rests on.

Drafting

list_templates · draft_document · review_draft

Six templates in correct Indian form: section 138 NI Act notice, general legal notice, reply to notice, Consumer Protection Act complaint, RTI application, general affidavit. Each carries its own procedural checklist and next steps. Amounts use Indian digit grouping (₹2,00,000, not ₹200,000).

Courts and jurisdiction

get_case_status · court_directory · determine_jurisdiction

Quick start

make install

make corpus

cp .env.example .env

Set ENABLE_AUTH=False in .env for local use, then:

make local

Download the free case-law corpus for the courts and years you need (one time, no API key, no charge). From your MCP client:

Run sync_case_law for the Supreme Court and Bombay High Court, 2015 to 2026.

Check what is covered any time with case_law_status.

Register it with Claude Code:

claude mcp add legal --transport http http://localhost:5001/mcp

Everything except case law, document storage and semantic search works with no API key and no database.

Optional capabilities

Install an embedding backend with uv pip install -e ".[voyage]" or ".[local]".

Try it

A cheque for ₹2,00,000 was dishonoured on 15 July 2026. What are my deadlines, draft the notice, and find Bombay High Court authority on territorial jurisdiction.

This chains compute_cheque_bounce_timeline → get_section → draft_document → search_case_law → verify_all_citations.

Architecture

legal_mcp_server/src/
├── mcp.py          LegalMCPServer; carries the advocate persona as FastMCP instructions
├── settings.py     Pydantic settings, extended with the legal configuration
├── domain/         Pure logic, no I/O — the most valuable and most testable code
│   ├── citations.py           Indian citation parsing (AIR, SCC, SCC OnLine, neutral)
│   ├── limitation.py          Limitation Act 1963 + special statutory periods
│   ├── holidays.py            Court closures and calendar confidence
│   ├── new_criminal_codes.py  IPC-BNS, CrPC-BNSS, Evidence-BSA concordance
│   └── clause_rules.py        Contract taxonomy and India-specific risk rules
├── sources/        External adapters, mockable, MCP-unaware
│   ├── indian_kanoon.py  Budget cap, cache, honest failure
│   ├── india_code.py     Bundled corpus with explicit coverage
│   ├── ecourts.py        Pluggable; no CAPTCHA automation
│   └── embeddings.py     voyage-law-2 | fastembed | disabled
├── storage/        legal_store.py owns its own pool and schema
├── templates/      Jinja2 documents + manifest.yaml
└── tools/          Thin MCP adapters over the above

Tool modules are deliberately thin. Legal logic lives in domain/, which needs no network or database to test — hence 98% coverage on limitation and citation parsing.

The server's instructions (in mcp.py) carry the practitioner brief the client reads on connect: cite everything, raise limitation unprompted, be direct about weaknesses, and say plainly that this is information rather than the advice of an engaged advocate.

Configuration

Every setting is documented in .env.example. The ones that change behaviour most:

Development

make test

make lint

make coverage

512 tests, 82% coverage, ruff and mypy clean.

make container

make deploy openshift NAMESPACE=legal-mcp

Run make corpus before building the image — the Containerfile copies data/, and without it every statutory lookup reports itself as unavailable. The compose file uses pgvector/pgvector:pg16; plain Postgres works but disables semantic search.

Further guides live in docs/: architecture, development, deployment, authentication and CI/CD.

Deliberate limits

• Not legal advice. This is legal information and drafting assistance. No lawyer-client relationship arises from it. For anything involving personal liberty, an imminent limitation date, a court appearance, or money you cannot afford to lose, engage an advocate.

• No CAPTCHA bypass. The official eCourts portal is CAPTCHA-gated. This server will not defeat that. Case status is manual-entry by default, or a licensed third-party API if you supply a key.

• It never sends anything. No filing, no service, no email. It drafts and organises; you decide what leaves the building.

• Coverage is bounded and stated. 19 Acts, 103 concordance mappings, 31 limitation rules — each tool reports what it does not cover rather than extrapolating.

• Your data stays yours. Matters and documents live in your PostgreSQL. Only search queries and citation checks leave the machine, plus chunk text to the embedding provider — which is why EMBEDDING_PROVIDER=local exists.

Sources

• redhat-data-and-ai/template-mcp-server — Apache 2.0

• Indian Supreme Court Judgments · Indian High Court Judgments - AWS Open Data, CC-BY-4.0, maintained by Dattam Labs. The default case-law source.

• Indian Kanoon API - optional paid alternative, not used by default

• India Code — authentic bare Act text

• civictech-India/Indian-Law-Penal-Code-Json — the open corpus make corpus fetches

• eCourts Services

License

Apache 2.0