LegalMike

LegalMike — MCP Server

MCP server for Dutch & European legal reasoning.

Overview

The LegalMike MCP Server gives your AI assistant Dutch and European legal reasoning with actual legal sources — legislation, case law, disciplinary and dispute-resolution decisions, EU and ECHR law, legal literature and parliamentary documents.

It is optimized to ground legal answers in retrieved text, so an assistant cites what it found rather than what it remembers.

  • Two tools: retrieve the passages themselves, or get a written answer with citations you can check.
  • The same corpora as the Golden Retriever, updated weekly.

Request access and find support at legalmike.ai or reach out to info@legalmike.ai.

Connect to LegalMike's MCP Server

You can connect LegalMike to Claude on Claude.ai, the Claude desktop app, and Claude Code.

Claude.ai

  1. Go to Settings → Connectors → Add Custom Connector
  2. Enter a name (e.g. "LegalMike") and the server URL: loading...
  3. Click Add — a browser window opens for you to log in with your LegalMike account
  4. In a conversation, click + → Connectors and enable LegalMike

Available on Free (1 integration), Pro, Max, Team, and Enterprise plans. See the Anthropic documentation for more details.

Claude Code

Run this command to add the server:

claude mcp add --transport http lmcp loading...

Add to .vscode/mcp.json in your project:

loading...

Then open Copilot Chat in Agent mode. Copilot will prompt you to authenticate with your LegalMike account via OAuth.

If OAuth isn't available, pass an API key in the Authorization header as a Bearer token.

You can add MCP apps to ChatGPT if you have a Pro, Plus, Team, Enterprise, or Edu account.

  1. Go to Settings → Apps & Connectors → Advanced settings and enable Developer mode
  2. Go back to Apps & Connectors and click Create
  3. Enter a name (e.g. "LegalMike") and the server URL: loading...
  4. Select "OAuth" as the connection mechanism and click Create
  5. In a new conversation, click + → More and select LegalMike

See the OpenAI documentation for more details.

Work in progress. Contact info@legalmike.ai for early access.

MCP clients with OAuth support

Use server URL loading... with "OAuth" as the connection mechanism. Your client handles the authorization flow automatically. 

loading...

API key (system-to-system)

For backend services, automated pipelines, or clients that don't support OAuth, pass an API key as a Bearer token in the Authorization header. No OAuth flow or client registration is needed. 

{
  "mcpServers": {
    "lmcp": {
      "url": "loading...",
      "headers": {
        "Authorization": "Bearer YOUR_API_TOKEN"
      }
    }
  }
}

API tokens are issued per organization with a configurable expiry. Contact info@legalmike.ai to request one for your organization.

Important: LegalMike provides legal information, not legal advice. Answers are grounded in authoritative sources but should be verified by a qualified legal professional. LegalMike does not store your questions or use them for training. By using the MCP server you agree to our terms of use.

Authentication

LegalMike requires authentication for all tool calls. Two methods are supported.

Method For How
OAuth 2.1 Individual users with a LegalMike account Your AI client handles the flow automatically (Authorization Code + PKCE). No keys to manage.
API Key (Bearer Token) System-to-system integrations, backend services LegalMike generates a long-lived token for your organization. Pass it as a Bearer token in the Authorization header. Contact info@legalmike.ai to request one.

Using an API key

Pass the token as a Bearer token in the Authorization header. No OAuth flow or client registration is needed. 

curl -X POST https://mcp.legalmike.ai/mcp \
  -H "Authorization: Bearer YOUR_API_TOKEN" \
  -H "Accept: application/json, text/event-stream" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","method":"initialize","params":{"protocolVersion":"2025-03-26","capabilities":{},"clientInfo":{"name":"my-app","version":"1.0"}},"id":1}'

Contact info@legalmike.ai to request an API key for your organization.

Client compatibility

Client OAuth 2.1 API Key
ChatGPT supported —
Claude supported —
Microsoft 365 Copilot supported supported
GitHub Copilot supported supported
Custom backend / API — supported

Available Tools

LegalMike's MCP server provides tools to search Dutch and European legal sources and synthesize grounded answers. Your AI assistant can call these tools directly — every claim in the response is backed by a citation to the original source. You don't need to learn the tools yourself; your AI client discovers and invokes them automatically based on your questions.

It is a filter, always. The sources you name are the sources searched and nothing else is. Omit the parameter and the server routes the question itself, which is the default and the right choice for most calls. Your profile's entitlements are applied on top and are not overridable.

Loading tools...

What Comes Back

Every call returns the answer as text and a structuredContent object, described by the tool's outputSchema — so you can validate it rather than reverse-engineer it. Your MCP client exposes the schema; it is also in GET /api/tools.

Branch on answer_type first. It is the one field always present, and it tells you which exit you are holding — because they carry different things. A clarification has no citations; a raw payload has no synthesized answer.

answer_type What you got
answer A synthesized, cited legal answer.
meta:… An answer about the assistant itself, not about law.
clarification A question back. clarification may carry structured options.
correction The question rests on a misconception; the text corrects it.
rejected Outside the legal domain.
non_legal Accepted, then found not to be a legal question after routing.
verbatim Literal statute or ruling text, in verbatim. No synthesis ran.
canned A fixed informational response; no research ran.
raw Retrieved passages per source, from legalmike_search.
error Only from legalmike_search — see below.

Alongside it: applied_settings (what the flow actually used, including sources_used), flow (the stages and sources actually traversed, in order), ecli_citations, and verbatim. New fields may be added — the schema allows them, so validate permissively.

When Something Goes Wrong

Failures follow the MCP specification's two channels. A problem with the request itself — unknown tool, malformed call — comes back as a JSON-RPC protocol error. Anything that happens while answering comes back as a tool execution error: isError: true with a message your assistant can act on, so it can rephrase or retry rather than simply failing.

Every failure carries a stable code. Which of the two tools you called decides whether that code reaches you, and the difference matters if you are writing a client:

Tool What you receive How to branch
legalmike_reason A tool execution error: isError: true and the message below as text. The code is not transmitted — it is recorded on the server’s trace, not in the response. Treat any isError as retryable and let the model read the message — it is written to tell the model what to try next.
legalmike_search A normal result whose payload carries answer_type: "error" with code and message, plus a sources array when the failure concerns specific sources. Branch on code. A caller parsing JSON cannot be handed an exception, so this is where the machine-readable contract is on the wire.
If you need codes, call legalmike_search. Attaching structured content to an isError result is not expressible in the server framework today, so the answering tool carries prose only. Messages are static and never interpolate anything, which makes matching on them stable in practice — with one deliberate exception: invalid_argument names the parameter, the value you sent and the values allowed, because that message is read first by a model that will retry, and “a parameter is invalid” gives it nothing to retry with.
Code Message (returned verbatim, in Dutch) What to do
no_evidence Ik heb geen bronnen gevonden om deze vraag op te baseren. Formuleer de vraag specifieker — noem bijvoorbeeld het rechtsgebied, een wetsartikel of een ECLI. Retry with a narrower question — name the field of law, an article, or an ECLI. Not retryable as-is.
sources_unavailable De juridische bronnen zijn tijdelijk niet bereikbaar. Probeer het over enkele minuten opnieuw. Transient. Retry after a short delay with backoff.
invalid_argument Een van de meegegeven parameters heeft een waarde die niet is toegestaan. The only code whose message is not static: it names the parameter, the value you sent and the values allowed, so the calling model can retry with a corrected one. Fix the argument; retrying unchanged fails identically.
retired_tool Deze tool bestaat niet meer. Gebruik 'legalmike_reason' voor een onderbouwd antwoord, of 'legalmike_search' voor losse bronnen. You called a tool that has been removed — in practice ask, which became these two. Like invalid_argument, the message is dynamic: it names the tool to call instead, so a model reading it can retry correctly. Same question, same parameters, new name.
upstream_timeout De zoekopdracht duurde te lang en is afgebroken. Een kortere of specifiekere vraag lukt vaak wel. Retry once with a shorter or more specific question. Repeated timeouts mean the question is too broad.
not_configured De server is niet volledig geconfigureerd. Neem contact op met de beheerder. Not retryable. A server-side configuration problem — escalate.
internal_error Er is een onverwachte fout opgetreden bij het beantwoorden van de vraag. Probeer het opnieuw; blijft het misgaan, meld het bij de beheerder. Retry once. If it persists, escalate with the timestamp; details are in the server trace, never in the response.
From legalmike_search an error is returned as a normal result like every other outcome, rather than as an exception — a caller that is parsing JSON cannot be handed one: 
{"answer_type": "error", "kind": "error", "code": "no_evidence", "message": "..."}
When the error concerns specific sources, a sources array names them. This payload arrives both as text and as structuredContent, and it satisfies the tool's declared output schema like every other exit.

Partial results. If a source is unreachable, the answer is still synthesized from the sources that responded and the server emits an MCP error-level log notification naming what was missing. An answer built from fewer sources than intended always says so.

Unexpected internal failures are deliberately opaque (Error calling tool 'legalmike_reason'). Details stay in the server's traces so nothing internal can reach a model context.

Legal Sources

The values below are what restrict_to_sources accepts, on both tools. The list you pass is a filter, not a hint — name three sources and exactly those three are searched. Omit the parameter (the default) and the server routes the question itself, which is the right choice for most calls. There is no exclude list: whatever you leave out is left out.

They are also the names returned in applied_settings.sources_used, so what you ask for and what you are told was searched use one vocabulary. Citations link directly to the original documents on their official platforms.

Product of LegalMike — ©