eu-reg-mcp

Let your AI assistant answer MiCAR and EU AI Act questions with citations instead of guesses — every finding flagged for human legal review.

Scope today: the ESMA interim MiCAR register (search, summary, change tracking), MiCAR white paper linting, and EU AI Act risk-tier classification. Other EU regulation is not covered.

Screenshot/GIF placeholder — to be added

Who it's for

• Crypto-asset compliance and legal teams looking up entities on the ESMA interim MiCAR register (white papers, authorised CASPs, non-compliant entities) and tracking what changes on it, including removals.
• Teams preparing MiCAR white papers who want a deterministic Annex I–III check of a draft that returns cited candidate findings.
• Teams assessing AI systems under the EU AI Act who want a first-pass risk-tier classification with pinpoint Art. 5 / Annex III / GPAI citations.

It does not replace legal review: its outputs are register facts or candidate flags for a human to assess (see Guardrails).

Try it now

There is no hosted playground. eu-reg-mcp is a Model Context Protocol (MCP) server, so you try it by adding it to your own MCP client — the Quick start has the configuration for Claude Code and Claude Desktop. Then ask your assistant:

• "Is <entity> on the MiCAR register, and what format is its white paper?"
• "What changed on the MiCAR register recently — any removals?"
• "Classify this AI system under the EU AI Act" (paste a system profile; call with an empty string to get the schema).

The first two work without extras; the third needs the classify extra.

Quick start

eu-reg-mcp is on PyPI and requires Python 3.13 or newer. With uv installed, uvx fetches and runs it.

Claude Code

claude mcp add eu-reg -- uvx eu-reg-mcp

The register tools need no API key. For all five tools, add the lint and classify extras instead:

claude mcp add eu-reg -- uvx --from 'eu-reg-mcp[lint,classify]' eu-reg-mcp

{
  "mcpServers": {
    "eu-reg": {
      "command": "uvx",
      "args": ["eu-reg-mcp"]
    }
  }
}

Without uv, run pip install 'eu-reg-mcp[lint,classify]' in a Python 3.13+ virtual environment and register its eu-reg-mcp by full path, e.g. claude mcp add eu-reg -- /path/to/.venv/bin/eu-reg-mcp; the environment need not be active. Install options and network use: docs/tools.md.

Example output

Synthetic data. Produced on 2026-10-14 by running this repo's register code (register.search_entries / register.recent_changes) against the synthetic test fixtures in tests/fixtures/, not the live register. The entity and its LEI are test data. Output shape is what a live call returns; values are not.

search_micar_register(query="skygate"):

{
  "snapshot_date": "2026-07-07",
  "total_matches": 1,
  "results": [
    {
      "register": "other-wp",
      "entity_name": "SKYGATE Network GmbH",
      "lei": "984500BBFE52FE449926",
      "member_state": "AT",
      "authority": "FMA",
      "whitepaper_url": "WWW.SKYGATETOKEN.AT",
      "format_class": "unspecified",
      "register_last_update": "30/12/2024"
    }
  ],
  "truncated": false,
  "notice": "Facts from the public ESMA interim MiCAR register (Art. 109 MiCAR) as mirrored by the MiCAR Register Observatory. No legal assessment of any entity; not legal advice."
}

micar_register_changes(change_type="removed") returns the same kind of cited record (snapshot date, register, entity) and notes that "the register itself does not state why" an entry was removed. With the extras installed, the lint and classify tools cite the provision behind each finding, for example Anhang I Teil A MiCAR (MiCAR Annex I, Part A), Art. 5(1)(f) AIA, Annex III(4)(a) AIA, and Art. 51 AIA for a general-purpose AI model with systemic risk. Sample output from synthetic inputs: docs/tools.md.

Tools and data sources

An MCP server that gives AI agents cited, review-gated grounding in EU regulation.

Five tools:

Parameters, data sources, caching and network requirements: docs/tools.md.

Why not just ask an LLM?

An ungrounded model can invent register entries or provisions. Here, register answers come from a dated snapshot of the public register, and lint/classification output comes from deterministic rules with citations. Review-gated means the output is a candidate flag pending human legal review — never a legal assessment or legal advice. Details: docs/guardrails.md.

Links

• PyPI: eu-reg-mcp, micar-whitepaper-linter, eu-ai-act-classifier
• Register data: micar-register-observatory (weekly snapshots of the public ESMA register, Art. 109 VO (EU) 2023/1114)
• Linter: micar-whitepaper-linter (36 rules mapped to Annex I–III) · Classifier: eu-ai-act-classifier — both MIT-licensed
• Development and tests: docs/development.md

Part of the EU regulation as code suite

By sebastianfoerste:

• awesome-eu-fintech-regulation
• contract-review-eval-harness
• legal-function-operating-system
• eu-ai-act-classifier
• micar-whitepaper-linter
• eu-reg-mcp (this repo)
• micar-register-observatory

Caveats

Register results are public-register facts with extraction artifacts possible; lint and classification outputs are candidate flags pending human legal review. Nothing here is legal advice. MIT licensed.