Legalize

Legalize — the MCP connector for consolidated legislation

What a law said on any past date, with the citation and the git commit behind it. Your assistant answers from the corpus instead of from memory.

Connect your AI client (Claude, ChatGPT, Cursor, VS Code, Gemini CLI…) to Legalize and ask about legislation in plain language. Every answer comes back with three things a model cannot invent:

• 📜 The text itself — the whole law or one whole article, never a truncated fragment.
• 🕰️ A date — not only the current wording, but the version that was on the books on the day you name.
• 🔗 A citation and a git SHA — the commit in the public corpus that holds those exact bytes, so the other side can check the quote.

One remote endpoint:

`

https://legalize.dev/mcp

`

Sign-in required. See it live at legalize.dev/mcp, which always shows the current coverage and allowance.

---

Quick connect

It is a remote server (Streamable HTTP, stateless). Nothing is installed: you hand the URL to your client, and the client walks you through the sign-in the first time it calls a tool.

One click, if yours is one of these:

The first three pre-fill the name and the address, then ask you to confirm it and sign in. ChatGPT opens its create-connector form — it takes no parameter for the server, so paste https://legalize.dev/mcp in yourself, with developer mode on under Advanced.

Claude.ai · Claude Desktop (Connectors)

Settings → Connectors → Add custom connector → paste https://legalize.dev/mcp.

Claude Code (CLI)

claude mcp add --transport http legalize https://legalize.dev/mcp

Cursor

Use the button above, or in ~/.cursor/mcp.json:

{

"mcpServers": { "legalize": { "url": "https://legalize.dev/mcp" } } }

VS Code (GitHub Copilot)

Use the button above, or in .vscode/mcp.json:

{

"servers": { "legalize": { "type": "http", "url": "https://legalize.dev/mcp" } } }

ChatGPT

Settings → Connectors → Advanced → turn on developer mode, then create a connector and paste https://legalize.dev/mcp. There is no install link for this one: ChatGPT only deep-links to apps listed in its own directory, and custom connectors sit behind developer mode, on Plus or Pro.

Gemini CLI and other clients

{

"mcpServers": { "legalize": { "httpUrl": "https://legalize.dev/mcp" } } }

Any client that speaks remote MCP works. If yours only speaks stdio, bridge it:

npx mcp-remote https://legalize.dev/mcp

---

The tools

Nothing here can change a law. The corpus is read-only to the connector: no tool alters a text, a version or a history, because those come from the git repositories and this server has no way in. The full descriptions the model reads, with every argument and its type, are published live by the running server at /mcp/tools.json.

The rest manage your own subscriptions — the one question reading the corpus cannot answer, which is *tell me when this changes*. They are the only tools that write anything, what they write is your account's own subscription, and they need a paid plan.

One rule, two ways to receive it. A webhook posts every change to a server you run; a digest lists a day of them in your inbox:

The digest goes to the account's own address and there is no argument for another one. Not an omission: a free-text recipient would turn a connector into a way to send mail to somebody else, signed by our domain, on a model's say-so.

Narrow the subscription or you will stop reading it. The filters compose as AND, a rule that could never match — an unknown country, a law id that is not in the corpus — is refused rather than stored, and a text filter reaches one language: laws are written in their own, so *proteccion de datos* finds the Spanish act and not the Portuguese *protecao de dados*.

The ones marked ✎ are declared to clients with readOnlyHint: false, so a client that confirms before a write will confirm before these. A paid plan is needed to create one; pausing and deleting never are, on any plan, because the account that most needs to turn its deliveries off is the one whose plan has just ended. On a plan without the feature, creating answers a structured feature_not_available error and creates nothing — and nothing is delivered either, to an endpoint or to an inbox, until it is back. Neither delivery is instant: webhooks are signed and batched daily (the format and the verifier), and a digest is one mail a day — on a day when nothing matched, no mail at all.

Every tool returns the same envelope — data, citation, url, source, note — and every result links back to its page on legalize.dev. A tool-level failure is a *successful* call whose data is a structured error naming what to do next, so the model can correct itself instead of guessing.

Examples (ask your assistant)

• *"What did article 348 bis of the Spanish Companies Act say in March 2023?"*
• *"Has the French Code du travail changed on this point since 2020? Show me the diff."*
• *"Which norms have amended this law, and when?"*
• *"How much of Latvian legislation does Legalize actually hold?"*
• *"Email me every morning when anything about data protection changes in Spain — in Spanish."*

Authentication

Sign-in required; there is no anonymous access. Every call is tied to an account, and an unauthenticated request gets 401 with the WWW-Authenticate challenge that starts the OAuth flow.

Your client discovers the authorization server from https://legalize.dev/.well-known/oauth-protected-resource, registers itself dynamically, and sends you to sign in once in a browser. After that it holds the token and you never see the handshake again. No API key is pasted anywhere.

Limits

There is a free monthly allowance and a per-minute burst guard. Neither number is written here on purpose — they would go stale the day either changes, and a README that quietly lies about a limit is worse than one that points at it. Both are stated, live, on legalize.dev/mcp and legalize.dev/pricing.

What does not change: past the monthly allowance the connector answers quota_exceeded and stops until the month rolls over — nothing is charged and nothing is cut off silently. Every country is included in every tier.

Before you quote it

Most of the corpus is consolidated: amendments are folded into the text, so what you read is the law as it stands. Some of it is not. A text published as enacted comes back flagged, with the norms that amend it named, and never as the law in force — but the flag is only useful if you read it before pasting the quote into a brief.

Three more limits worth knowing:

• A date resolves to what was published on or before it, not to what was in force.
• Article boundaries are derived by matching headings in the text, not published as structure by the source. Quote the anchor, and check the boundary before relying on it.
• Counts are of what Legalize has ingested, not of what the official gazette has published. A gap in the corpus is not evidence that a norm does not exist.

A body is never truncated: a law too large to send whole comes back as its table of contents with instructions for asking again, because half an article reads exactly like a whole one.

For developers

Transport is Streamable HTTP, stateless. Every JSON-RPC method needs a token, initialize and tools/list included. So the first thing to check is the catalogue, which is served beside the endpoint as a plain GET and needs nothing:

curl -s https://legalize.dev/mcp/tools.json | jq '.tools[] | {name, readOnly}'

That is what a directory reads, and it is where readOnly per tool is published. The JSON-RPC endpoint itself answers 401, and that 401 is the handshake — its WWW-Authenticate header is what tells a client where the authorization server is:

curl -si -X POST https://legalize.dev/mcp \

-H 'Content-Type: application/json' \ -H 'Accept: application/json, text/event-stream' \ -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"lis