openlaw-mcp

by LegalAI Space

A free, open-source Model Context Protocol (MCP) server that exposes UK and EU open law — judgments, statutes, regulations, regulator notices — to host LLMs such as Claude, ChatGPT, and Cursor.

Where general-purpose chatbots invent citations, openlaw-mcp fetches the real text from official open-data sources (The National Archives' Find Case Law, legislation.gov.uk, EUR-Lex, HUDOC, ICO, HMRC) and returns it with full provenance and licence attribution. Every tool result is cite-back ready.

Sample prompts

Once the server is configured, paste any of these into your host LLM. They are UK-focused and chosen to exercise the tools where MCP grounding makes the biggest difference — exact statute text, point-in-time status, judgment paragraph quotes, regulator deep links, and ICO / HMRC material that a plain web search cannot reliably return verbatim.

Statute lookups (lookup_statute, legislation_toc)

• "Quote subsection (1) of section 1 of the UK Bribery Act 2010 verbatim, with a link to the legislation.gov.uk source."
• "What is the current in-force text of section 1 of the Online Safety Act 2023 (c.50) on legislation.gov.uk? If the section has not been commenced, say so and cite the source you used."
• "What did section 2 of the Data Protection Act 2018 say as in force on 25 May 2018? Use legislation.gov.uk's point-in-time view."
• "Give me the table of contents for the Procurement Act 2023 — parts, chapters, and section numbers."
• "Compare the wording of section 1 of the Computer Misuse Act 1990 as enacted vs as currently in force, and cite both versions on legislation.gov.uk."

Case law (search_caselaw, fetch_judgment, grep_judgment, citator_lite)

• "Quote paragraph 1 of R (on the application of Miller) v The Prime Minister [2019] UKSC 41 verbatim, with the Find Case Law URL."
• "Find UK Supreme Court judgments from 2024 onwards that consider Article 8 of the ECHR. Return neutral citations with Find Case Law links and a one-line summary of each."
• "In Uber BV v Aslam [2021] UKSC 5, find every paragraph that mentions the word 'control' and quote each one in full."
• "List judgments published on Find Case Law that cite Pepper v Hart [1992] UKHL 3. Return neutral citations and links — no treatment classification needed."
• "Search Find Case Law for First-tier Tribunal (Tax Chamber) decisions on IR35 / off-payroll working handed down in 2023, and return the first 10 hits."

Regulator material (search_fca_handbook, fetch_fca_notice, search_ico_guidance, fetch_hmrc_manual)

• "Give me the official handbook.fca.org.uk deep link for SYSC 4.1.1R."
• "Resolve these FCA rule references into deep links: SYSC 6.1.1R, COBS 9.2.1R, CASS 7.13.3R."
• "Find recent ICO enforcement actions (monetary penalty notices and enforcement notices) issued in 2025 — return the title, date, and source URL for each."
• "From the HMRC Business Income Manual, give me the section on the badges of trade (BIM20205). Return the bodyText and the gov.uk URL."
• "From the HMRC Employment Status Manual (ESM), summarise the contents page (just the navigation tree, not the bodies)."

EU material (fetch_eurlex)

• "Fetch CELEX 32016R0679 (the GDPR) and return Article 5 in full, with the EUR-Lex source URL."
• "Fetch CELEX 32024R1689 (the EU AI Act) and tell me which articles deal with prohibited AI practices."

ECHR / Strasbourg (fetch_hudoc)

• "Pull the full English text of the ECHR judgment in Bărbulescu v Romania (Application 61496/08) and quote the operative paragraphs."
• "Search HUDOC for Grand Chamber judgments where the United Kingdom was the respondent state between 2020 and 2024."

Citation hygiene (format_oscola)

• "Format the following bundle of UK cases in OSCOLA, one per line: Donoghue v Stevenson [1932] AC 562; Pepper v Hart [1992] UKHL 3; R (Miller) v The Prime Minister [2019] UKSC 41."
• "Convert this sloppy reference into OSCOLA and produce a working Find Case Law URL: 'Uber BV v Aslam, 2021 Supreme Court, UKSC 5'."

When MCP shines (and when it doesn't)

• Strongest payoff: "verbatim text of section X" prompts, point-in-time statute lookups, FCA Handbook deep-link resolution, and any request that needs current in-force wording or a recent ICO/HMRC publication. Generic web search either guesses or gets blocked by upstream JS challenges (notably legislation.gov.uk's AWS WAF); the MCP server is allow-listed and returns parsed, cite-ready XML.
• Marginal payoff: quoting a famous paragraph from a landmark judgment, or formatting a textbook citation. The host LLM often gets these from training data alone — MCP still adds provenance and reduces hallucination risk, but the wording would have been correct anyway.
• No payoff: purely conceptual questions ("explain the difference between the FCA and the PRA"). These should be answered without calling any tool; if your host LLM fires MCP tools here, treat it as an over-eager router rather than a useful grounding.

Tools

Resources

Clients that support MCP resources can also read:

How tools and resources work technically

Common runtime flow (all tools)

1. MCP client sends tool name + JSON arguments (inputSchema validated by zod).
2. Server handler calls either:
   • a source client (src/sources/*.ts) that performs HTTP fetches via shared Fetcher, or
   • a local computation path (FCA link resolvers, citation formatter) with no network call.
3. Handler normalizes/parses upstream data into stable data fields.
4. Server returns an envelope: data, provenance (source, source_url, fetched_at, content_hash, cache_hit, licence metadata), and learn_more; some tools also include note.
5. On upstream errors, tools return isError: true plus a canonical authoritative URL.

Per-tool technical behavior

• search_caselaw: receives search filters (query, optional court/date/page), fetches TNA Atom feed, parses entries into normalized judgment refs, returns paged hits + provenance.
• fetch_judgment: receives neutral citation or TNA URI, resolves to /data.xml, parses Akoma Ntoso into meta, header, paragraphs, returns full structured judgment + provenance.
• grep_judgment: same fetch/parsing as fetch_judgment, then runs JS regex across parsed paragraph text, returns match list/snippets + provenance.
• lookup_statute: receives legislation type/year/number/section (+ optional as_of), fetches CLML section XML, extracts title/text/version metadata/extent/prospective flags, returns normalized section document + provenance.
• legislation_toc: receives legislation type/year/number, fetches CLML contents XML, flattens parts/chapters/sections into entries, returns TOC + provenance.
• fetch_eurlex: receives CELEX, validates/normalizes it, blocks out-of-scope national sectors (7/8), fetches from Cellar with content negotiation (XHTML then HTML fallback), returns raw body + detected media type + provenance.
• fetch_hudoc: receives either itemid or search query. Search path fetches JSON results and suppresses out-of-policy records. Item path fetches metadata then body only when language/doctype/translation guard passes. Returns records/body + provenance + suppression note when applicable.
• search_fca_handbook: receives parsed handbook reference inputs, computes deterministic deep link locally, returns metadata + local provenance (urn:openlaw-mcp:local) without FCA content fetch.
• fetch_fca_notice: receives known notice slug (+ category), computes deterministic FCA deep link locally, returns metadata + local provenance; listing path intentionally errors (scraping disabled).
• search_ico_guidance: receives optional type/sector/date/page filters, posts to ICO /api/search, shapes response into enforcement list items, returns paged results + provenance.
• fetch_hmrc_manual: receives manual slug (+ optional section slug), fetches GOV.UK Content API JSON, returns either manual navigation tree or section body (bodyHtml + bodyText) + provenance.
• format_oscola: receives citation text + formatting options, parses and formats locally (OSCOLA/Bluebook), optionally emits derived Find Case Law URL for neutral citations, returns formatted output + local provenance.
• citator_lite: receives neutral citation, searches TNA Atom with quoted query to find citing judgments, trims to limit, returns citing list + disclaimer + provenance.

Per-resource technical behavior

• judgment://.../header: MCP resource vars (slug, year, number) are validated/parsed, server fetches TNA judgment XML once, extracts plain-text header, returns text/plain.
• judgment://.../index: same fetch path, maps parsed paragraphs to { eId, num, preview }, returns application/json.
• judgment://.../para/{eId}: same fetch path, selects one paragraph by exact Akoma Ntoso eId, returns text/plain ([num] text when num exists).
• statute://.../section/{section}: validates resource vars, fetches legislation.gov.uk section XML via legislation client, returns parsed section document JSON.

Provenance

Every tool response carries a provenance block:

{
  "data": { },
  "provenance": {
    "source": "Find Case Law",
    "source_url": "https://caselaw.nationalarchives.gov.uk/uksc/2024/12/data.xml",
    "fetched_at": "2026-05-12T12:00:00Z",
    "content_hash": "sha256-...",
    "cache_hit": false,
    "licence": "Open Justice Licence v2.0",
    "licence_url": "https://caselaw.nationalarchives.gov.uk/open-justice-licence",
    "attribution": "Contains information licensed under the Open Justice — Licence v2.0"
  },
  "learn_more": "https://legalaispace.com"
}

If you redistribute the content downstream, preserve the attribution.

Cache

By default, GET responses are cached in ~/.cache/openlaw-mcp/http-cache.sqlite with source-specific TTLs. Delete the cache directory to clear stored responses.

Multi-tenant deployments: The cache is shared across all callers of a given process. The TNA Find Case Law content is licensed under the Open Justice Licence v2.0, whose "computational analysis" carve-out does not permit a persistent cross-user retrieval index. If you are running the HTTP transport for more than one end-user, either set OPENLAW_CACHE_DIR=off or give each tenant its own ephemeral cache directory. Single-user stdio usage (npx @legalaispace/openlaw-mcp) is unaffected.

Takedown

If you are a rights-holder and believe content fetched by this server has been redistributed in error, please email daman@legalaispace.com. We will investigate within five working days.

Development

npm install
npm run dev:stdio          # run the stdio server locally
npm run dev:http           # run the Streamable HTTP server with watch
npm test                   # unit tests, MCP protocol tests, and the ICO fixture test
npm run lint && npm run typecheck
npm run build
npm pack --dry-run         # verify npm package contents; runs release gates

Install

Three ways to use openlaw-mcp:

1. Free hosted endpoint — fastest path, zero install, point your client at our URL.
2. Local stdio — runs inside Claude Desktop / Cursor / Claude Code via npx.
3. Self-hosted HTTP — node dist/http.js after npm run build.

1. Free hosted Streamable HTTP endpoint

A community endpoint is available at https://openlawmcp.legalaispace.com/v1/mcp. No API key, no signup, no rate-limit gate beyond the per-IP throttle. Best-effort uptime — no SLA, may be rate-limited, throttled, or migrated as load grows.

Codex CLI

codex mcp add openlaw --url https://openlawmcp.legalaispace.com/v1/mcp

Claude Code CLI

claude mcp add --transport http openlaw https://openlawmcp.legalaispace.com/v1/mcp

Claude Desktop / Cursor (config file)

{
  "mcpServers": {
    "openlaw": {
      "url": "https://openlawmcp.legalaispace.com/v1/mcp"
    }
  }
}

Health check: curl https://openlawmcp.legalaispace.com/healthz should return {"status":"ok","service":"openlaw-mcp"}.

2. Claude Desktop / Cursor / Claude Code (stdio)

Runs locally — no dependency on the hosted endpoint.

{
  "mcpServers": {
    "openlaw": {
      "command": "npx",
      "args": ["-y", "@legalaispace/openlaw-mcp"]
    }
  }
}

3. Self-host the HTTP transport

npm install
npm run build
node dist/http.js   # listens on PORT (default 3000), mount /v1/mcp

Behind a reverse proxy (nginx, Cloudflare, etc.), set TRUST_PROXY to the number of proxy hops so express-rate-limit keys by the real client IP — see Configuration.

Configuration

All configuration is via environment variables.

Limitations

• No BAILII. BAILII's terms prohibit programmatic re-use; the client blocks all outbound requests to it.
• No FCA scraping or full text. FCA Handbook and Notices are link-only / metadata-only by design. Known slugs and references are converted into deep links; the server does not enumerate FCA notices or fetch FCA page/PDF bodies without written permission.
• HUDOC scope. Only English/French Court-authored judgments. Third-party translations are excluded.
• Find Case Law coverage is E&W courts and tribunals from 2001 onwards. Older judgments are out of scope.
• EUR-Lex Member State legislation reached via N-Lex is excluded; only EU-authored content is fetched.
• legislation.gov.uk WAF. www.legislation.gov.uk is fronted by an AWS WAF that issues a JavaScript challenge (HTTP 202 Accepted with x-amzn-waf-action: challenge and an empty body) to clients whose IPs it considers suspicious — typically datacenter egress. From a residential network or an allow-listed deployment the API serves XML normally. Affected calls (lookup_statute, legislation_toc) surface a clear error and retry with backoff; we do not bypass the challenge by fingerprint-spoofing.
• Caselaw is not training data. openlaw-mcp fetches on demand for individual user queries; it does not maintain a persistent cross-user index of fetched content.

Troubleshooting

lookup_statute / legislation_toc fail with an AWS WAF challenge

The error you see in your host LLM will read something like:

upstream responded with an AWS WAF JavaScript challenge (HTTP 202) on every attempt

www.legislation.gov.uk is fronted by an AWS WAF that issues a JavaScript challenge to clients whose IPs it considers suspicious — typically datacenter egress, some VPNs, and an evolving list of residential prefixes. The server detects the challenge (HTTP 202, x-amzn-waf-action: challenge, empty body) and surfaces a structured error rather than spoof fingerprints to bypass it. See Limitations for the design rationale.

What to do:

• Run from a residential or allow-listed network, or point your host LLM at the free hosted endpoint, which runs from an allow-listed IP.
• The legislation tools (lookup_statute, legislation_toc) are the only tools affected by the legislation.gov.uk WAF. Caselaw (search_caselaw, fetch_judgment, grep_judgment), ICO, HMRC, EUR-Lex, HUDOC, FCA Handbook, and OSCOLA formatting all use independent upstreams and continue to work.
• Quick sanity check from your shell:
  Copy

  curl -s -o /dev/null -w "%{http_code}\n" https://www.legislation.gov.uk/ukpga/2010/23/section/1/data.xml

  200 → the API is reachable from your network. 202 → WAF is challenging your IP.

npx -y @legalaispace/openlaw-mcp fails to start in Claude Desktop / Cursor / Claude Code

• Confirm node --version is >= 22 (engines.node in package.json). On Windows, install Node via nvm-windows or fnm; on Linux, prefer your distribution's Node 22+ package or Volta.
• First-run npx downloads the package; subsequent runs are cached. If your host LLM kills the spawn before the download completes, run npm install -g @legalaispace/openlaw-mcp once and switch the config to { "command": "openlaw-mcp" }.

Cache acting up

Cached responses live in ~/.cache/openlaw-mcp/http-cache.sqlite (override with OPENLAW_CACHE_DIR; disable with OPENLAW_DISABLE_CACHE=1). Delete the file to force a fresh fetch from upstreams. See Cache for the multi-tenant guidance.

Circuit breaker tripped

The shared fetcher trips a per-host circuit breaker on repeated 5xx / network errors and cools down for 30 seconds before retrying. If you see "circuit open" errors, wait, then retry — and check the upstream's own status page.

License and attribution

Source code is Apache-2.0; see LICENSE. Upstream-data attributions and prior-art credit (notably paulieb89/uk-legal-mcp) are in ATTRIBUTIONS.md.

Development assisted by AI and Daman Kaur.

The information returned by this server is for information and research only and does not constitute legal advice.