MCP server

Agent quick start: discover, choose, call, check

This page is a self-contained reference for all nine tools. Step IDs, option IDs and types below match the live tools/list schema checked 26 October 2026. Never invent IDs or fill in unknown facts. Refresh the schema after connecting to detect future changes.

1. Read this reference or call tools/list to discover parameters. A normal MCP client handles its protocol handshake; legacy clients use initialize, notifications/initialized, then tools/list.
2. Choose the tool below. If you cannot identify a provision number, use search_text, then read the returned provision in full.
3. Call tools/call with the tool name and an arguments object. Parameter names, types and allowed IDs are case-sensitive.
4. Check JSON-RPC errors and result.isError before using content. Preserve the source, date, context and limitations when reporting an answer.

Direct schema discovery, no external inspector:

curl -X POST https://mcp.euaiactlaw.com/mcp \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}'

Tool-call envelope:

{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "tools/call",
  "params": {
    "name": "get_article",
    "arguments": {
      "number": "50"
    }
  }
}

Tools (9: eight read-only lookups, plus send_feedback)

“Required” is a schema requirement, not proof the supplied facts are sufficient for a sound assessment. Example arguments below fit the contract; they are not legal advice or automatic submissions.

• get_article: Read a known article. If its number is unknown, search first.
• get_recital: Read a known numbered recital.
• get_annex: Read a known annex, using its Roman numeral.
• get_definition: Find an Article 3 definition by term.
• search_text: Find passages when a topic or phrase is known but provision numbers are not.
• get_obligations: List obligations for one known role; call separately for each role.
• get_deadlines: Inspect the site's application timeline at a date.
• send_feedback: Explicitly submit a problem, gap or useful experience to the maintainer. This writes data.
• assess_compliance: Apply the interactive checker to known scenario facts; answer relevant route questions.

{
  "number": "50"
}

{
  "number": 1
}

{
  "number": "III"
}

{
  "term": "provider"
}

{
  "query": "recruitment",
  "scope": "all",
  "limit": 5
}

{
  "role": "deployer",
  "fields": "compact",
  "limit": 10,
  "offset": 0
}

{
  "date": "2027-06-01"
}

{
  "category": "content_gap",
  "message": "Describe the missing material and the source that supports it.",
  "tool": "get_article",
  "client": "your-client-name"
}

{
  "E1": [
    "deployer"
  ],
  "E2": [
    "none"
  ],
  "S1": [
    "established_eu"
  ],
  "R2": [
    "none"
  ],
  "R3": [
    "none"
  ],
  "R4": [
    "none"
  ],
  "R5": "no",
  "HR1": [
    "none"
  ],
  "HR2": [
    "none"
  ],
  "HR4": [
    "employment"
  ],
  "HR5": "yes"
}

assess_compliance: every step and valid option

The IDs below come directly from the tool schema, not from page headings or guessed numbers. There are 15 steps: E1-E3, S1, R1-R5 and HR1-HR6. Only E1 is technically required. All other keys are optional in the schema, but the route may need their facts. Omission is not an “unknown” answer: missing arrays can act like no selection. In particular, missing S1 produces out-of-scope even when other supplied facts generate a high-risk classification and obligations.

Before calling: collect the relevant facts. Multi-select steps take arrays of exact option IDs; HR3, HR5 and R5 take one string. Do not combine none with positive options. Do not invent a negative answer when facts are unknown; ask for the facts or report that assessment is incomplete. The server does not validate that every question on the route was answered.

The route descriptions below document the current checker implementation, not an independent legal conclusion. Use route, terminal and statuses to audit what happened. Human questionnaire: interactive compliance checker.

Responses, errors and coverage limits

• Read human-readable blocks in result.content. The assessment also has result.structuredContent and a checker resource link. Every tool declares an outputSchema in tools/list; responses also carry human-readable text blocks, and assess_compliance returns a structuredContent object matching its outputSchema. The presence of an output schema alone is not proof that every runtime response uses one uniform structured envelope.
• Check result.isError. A missing provision can return HTTP200 with a tool error instead of transport failure. Observed: get_article({"number":"999"}) returned isError:true and “No article numbered "999" found. Use search_text to find the right provision.” Search by topic and use returned references; never fabricate a number.
• Handle top-level JSON-RPC errors separately from tool errors. A successful HTTP response alone is not successful retrieval. Exact malformed-request and rate-limit envelopes are not specified here; do not retry indefinitely or guess thresholds.
• The assessment may return an out-of-scope terminal alongside classification and obligation data derived from raw answers. This is especially possible with omitted S1. Do not present such an internally mixed result as a final conclusion. Resolve missing facts and inspect the route, or state the limitation.
• Coverage is the site's English AI Act dataset and its amendment/version labels, not all EU/national law, case law or exhaustive legal advice. A matching passage or obligation list does not prove applicability, completeness or a correct legal conclusion. Verify the cited official text and dates.
• Respect modest automated request rates. Do not use send_feedback as a query channel, status poll or support conversation; it records a one-way message.

Worked agent workflow: recruitment tool

Fictional facts, real bounded read-only call checked 26 October 2026: an EU-established deployer uses an employment system, answers the significant-risk question yes, and has confirmed the listed negative answers. Those facts are assumptions for the example, never defaults to infer for a real user.

1. Find candidate provisions with search_text({"query":"recruitment","scope":"all","limit":5}). Follow returned labels and permalinks using get_article/get_annex; snippets are discovery, not full evidence.
2. Collect all relevant questionnaire facts. Call assess_compliance with the complete example arguments shown above.
3. Inspect the terminal and route. The observed structured summary was:

{
  "terminal": "RESULT",
  "route": [
    "E1",
    "E2",
    "HR1",
    "HR2",
    "HR4",
    "HR5",
    "S1",
    "R2",
    "R3",
    "R4",
    "R5"
  ],
  "roles": [
    "deployer"
  ],
  "classification": [
    "high-risk (Annex III)"
  ],
  "obligation_count": 11
}

This is the checker's result, not independent legal verification. Review the returned obligations, their legal bases and dates against the full source. Compare role obligations separately with get_obligations if useful; do not treat its unfiltered inventory as automatically applicable.

For an exact quotation rather than assessment, call get_article directly. For a date-specific timeline, call get_deadlines with an explicit ISO date. Submit feedback only when intended; it is not needed for any lookup.