# EU AI Act MCP Server

[![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE)
[![npm version](https://img.shields.io/npm/v/@lexbeam-software/eu-ai-act-mcp)](https://www.npmjs.com/package/@lexbeam-software/eu-ai-act-mcp)
[Connect to Claude](https://claude.ai/directory/lexbeam-eu-ai-act) (Anthropic's connectors directory)
[Smithery listing](https://smithery.ai/servers/lexbeam-software/eu-ai-act)
[![Test](https://github.com/lexbeam-software/eu-ai-act-mcp/actions/workflows/test.yml/badge.svg)](https://github.com/lexbeam-software/eu-ai-act-mcp/actions/workflows/test.yml)

An open-source [Model Context Protocol](https://modelcontextprotocol.io) (MCP) server that gives LLMs structured intelligence about the EU AI Act (Regulation (EU) 2024/1689, as amended by the Digital Omnibus, Regulation (EU) 2026/1744).

Built by [Lexbeam Software](https://lexbeam.com) - an agentic AI implementation boutique for regulated workflows.

## Listed use, not sector

Annex III regulates uses, not sectors. Set `signals.domain` to an Annex III area only when
the system itself performs a use listed there; a system that merely operates in that sector
gets `"other"`. Two calls to `euaiact_classify_system`, both for systems in the employment sector:

**In the sector, not a listed use.** Returns `limited` (Art. 50(1) transparency).

```json
{
  "description": "An HR chatbot answers employees' questions about the holiday policy and where to find the travel expense form.",
  "signals": { "domain": "other", "interacts_with_natural_persons": true }
}
```

**A listed use.** Returns `high-risk` (Annex III(4)).

```json
{
  "description": "Our ATS screens incoming CVs against the job spec and auto-rejects anyone missing the right keywords before a recruiter looks.",
  "signals": { "domain": "employment" }
}
```

The classifier trusts the signal: the same chatbot sent with `"domain": "employment"` comes
back `high-risk`. Leave `domain` out when the description does not tell you which it is.

## What's new in 1.6.0

Version 1.6.0 corrects what a calling agent is told, and measures what it then does.

- **Listed uses, not sectors:** `signals.domain` names an Annex III area only when the
  system itself performs a use Annex III lists for it. Until now the field read as the
  sector a system operates in, and an agent that filled it faithfully got a school
  timetable, a payroll check and a court budgeting tool back as high-risk. Measured with
  one agent model on 358 descriptions: 46 of 106 non-regulated systems were wrongly
  regulated under the old tool definition and none under the new one, while recognition
  of regulated systems held (181 and 179 of 252). See [evals/front-door](evals/front-door).
- **Signals first:** the tool description tells the agent to derive the structured
  signals from the user's description itself. Free text alone is only keyword-matched.
- **A callable assessment:** the `euaiact_assess_system` description states the fact
  shape and a useful minimum profile.
- **Releases with two human steps and no token:** a GitHub release stages the package on
  npm over trusted publishing, and it becomes public only when a maintainer approves it
  with 2FA. A weekly job checks every lexbeam.com link the server publishes, and a version
  bump no longer moves a pinned hash.

Full release history: [CHANGELOG.md](CHANGELOG.md).

## Quick Start

### Claude, one click

The hosted server is listed in Anthropic's connectors directory as a community connector:
[claude.ai/directory/lexbeam-eu-ai-act](https://claude.ai/directory/lexbeam-eu-ai-act).
Open the listing and connect; it works in claude.ai, the Claude apps and Cowork. No account with Lexbeam and no
API key are needed. The same server answers at `https://mcp.lexbeam.com/mcp` for any other MCP client (below).

### npx (no install)

```bash
npx -y @lexbeam-software/eu-ai-act-mcp
```

### Claude Desktop

Add to `claude_desktop_config.json`:

```json
{
  "mcpServers": {
    "eu-ai-act": {
      "command": "npx",
      "args": ["-y", "@lexbeam-software/eu-ai-act-mcp"]
    }
  }
}
```

### Smithery

```bash
npx -y @smithery/cli@latest mcp add lexbeam-software/eu-ai-act
```

Direct MCP endpoint: `https://mcp.lexbeam.com/mcp` (health check at `/health`). Open, no auth required. The
front page at [mcp.lexbeam.com](https://mcp.lexbeam.com/) has the tools and how to connect, and `/llms.txt` the same
facts as plain text.

The Smithery-hosted endpoint `https://eu-ai-act--lexbeam-software.run.tools` requires Smithery authentication and returns 401 without it.

### From source

```bash
git clone https://github.com/lexbeam-software/eu-ai-act-mcp.git
cd eu-ai-act-mcp
npm install
npm run build
npm start        # stdio transport
npm run start:http  # streamable HTTP (for Smithery/Railway)
```

## First successful calls

For the hosted server, add `https://mcp.lexbeam.com/mcp` as a remote Streamable HTTP
MCP server in your client. It requires no API key. The package also supports the local
stdio command above. Send the negotiated `MCP-Protocol-Version` header on subsequent
HTTP requests; browser clients can preflight that header.

Start with `euaiact_get_article` and `{"article":"Artikel 4"}` to retrieve the
existing Article 4 summary and its official source link. `4`, `Art. 4`, and
`Article 4` are equivalent whole-article lookups. Summaries are not statutory text.

`euaiact_answer_question` searches curated FAQs. For example,
`{"question":"Welche Pflichten zur KI-Kompetenz gelten nach Artikel 4?"}` routes to
the literacy FAQ. Answers remain in English. Check `match_status`: `matched`
identifies the accepted `matched_question`; `no_match` contains guidance instead
of a FAQ answer and may identify a rejected `candidate_question`. Confidence is
match confidence, not a probability that a legal conclusion is correct.

For the structured assessment, use the complete, syntactically valid
[minimum profile](examples/assessment-minimal.json) as the arguments to
`euaiact_assess_system`. It is synthetic and intentionally leaves unconfirmed facts
out. Replace its assertions only with facts the user supplied; do not copy the
example's answers into a real assessment. Read `missing_facts`, `legal_classification.limitations`, and
`recommended_next_calls`, even when the status is `determined`. Ask the relevant
questions, add fact objects with unique IDs, and call `euaiact_assess_system` again
to update all three result blocks. A minimal profile is not a complete assessment.

For the simpler classifier, `missing_signals` includes unprovided optional context.
It does not mean all listed fields must be collected before reporting a matched
route. Use `next_questions` for follow-up guidance. A matched high-risk route does
not establish that every prohibited-practice or other route was ruled out.

All tools reject unknown argument names, including nested objects, consistently
with their advertised closed input schemas. Correct a misspelled field rather
than retrying it silently. Omitted optional fields and documented defaults work
as before.

## Tools

| Tool | Description |
|------|-------------|
| `euaiact_classify_system` | Classify an AI system's risk level (prohibited / high-risk / limited / minimal) from structured signals that the calling agent derives from the user's description; free text alone is a keyword fallback. Returns matched signals, missing signals, and follow-up questions. |
| `euaiact_check_deadlines` | Implementation milestones with days remaining, `next_milestone` shortcut, `only_upcoming` filter, and the enacted Digital Omnibus (Regulation (EU) 2026/1744) status. |
| `euaiact_get_obligations` | Specific compliance obligations by role (provider/deployer) and risk level, including GPAI (Art. 51-56) and universal AI literacy (Art. 4). |
| `euaiact_answer_question` | Keyword FAQ search (lexical matching with stopword filtering, tie handling and abstention) across 24 curated EU AI Act questions; echoes your question and names the matched entry. |
| `euaiact_calculate_penalty` | Calculate maximum fines by violation type, turnover, SME status (Art. 99(6)) and SMC status (Art. 99(6a), tiers 99(4)-(5) only), with a comparative non-SME vs SME block. |
| `euaiact_get_article` | Retrieve an operational summary and EUR-Lex URL for a specific article. Covers 28 curated articles between Art. 3 and Art. 113 (including the new Art. 4a), not the full act. |
| `euaiact_check_gpai_systemic_risk` | Check whether a GPAI model crosses the 10²⁵ FLOPs threshold and return Art. 53 + Art. 55 obligations plus the Art. 52 notification duty. |
| `euaiact_assess_art6_3_exception` | Walk through the Art. 6(3) "no significant risk" exception with explicit profiling block and Art. 6(4) / Art. 49(2) reminders. |
| `euaiact_annex_iv_checklist` | Return all nine Annex IV technical-documentation items, optionally as a markdown checklist. |
| `euaiact_assess_system` | Assess a normalized system profile with separate legal-classification, qualitative-impact, and evidence-readiness blocks. Sparse decisive facts fail closed; every finding carries fact IDs and official provenance. |

## Resources

| URI | Description |
|-----|-------------|
| `euaiact://timeline` | Key implementation milestones of the EU AI Act. |
| `euaiact://risk-levels` | Overview of the four risk categories. |
| `euaiact://annex/iii` | Full Annex III high-risk AI categories (1-8) with descriptions, examples, and article references. |
| `euaiact://annex/iv` | Full Annex IV technical-documentation items (1-9). |
| `euaiact://omnibus` | The Digital Omnibus on AI (Regulation (EU) 2026/1744) as enacted: amended dates, deltas, and source status per item. |

## Prompts

- `classify-my-system` - guided classification using `euaiact_classify_system` with signal inference
- `compliance-checklist` - risk-level + role obligations checklist, including Annex IV for high-risk
- `penalty-risk-assessment` - penalty calculation with SME comparative
- `ground-citation` - retrieve an operational summary + EUR-Lex URL, then verify definitive wording in the official source before quoting

## Knowledge Base

Curated, structured data covering:

- **8 Annex III high-risk categories** with keyword matching and examples
- **10 prohibited AI practices**: Art. 5(1)(a)-(h), plus (ba) and (bb) applying from 2 December 2026
- **Art. 6(3) exception conditions** with the profiling block rule
- **Art. 50 transparency triggers** (chatbots, deepfakes, emotion recognition, machine-readable marking)
- **8 implementation milestones** with dynamic days-remaining calculation
- **Digital Omnibus (enacted)** status and impact assessment
- **Provider obligations** (13 for high-risk, 8 for GPAI including Art. 53 + Art. 55)
- **Deployer obligations** (9 for high-risk)
- **Limited-risk transparency obligations** (4 under Art. 50)
- **Universal AI literacy** (Art. 4)
- **Penalty framework** with SME protection logic (Art. 99)
- **25 FAQ entries** with article references
- **28 article summaries** with EUR-Lex URLs to the consolidated text
- **Annex IV (9 documentation items)** *(new in 1.1.0)*

Load-bearing dates, thresholds, amounts and exceptions are checked on every build by a claim matrix (`test-claims.mjs`) against a pinned, hash-verified copy of the consolidated act (`law/`). Coverage is the matrix, not a blanket claim.

## Regulatory Accuracy

This server tracks the current state of the EU AI Act (Regulation 2024/1689) **as amended by the Digital Omnibus on AI**, Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026 and in force from 27 July 2026. The amended application dates are served as operative law. Article-level wording was verified against the enacted OJ text on 26 and 27 July 2026, delta by delta against the 43 numbered amendments in Article 1 of the amending act. No Omnibus delta was left unresolved in that reconciliation. The Art. 49 registration duty for self-assessed not-high-risk systems, previously carried as unresolved, SURVIVES: the enacted act does not amend Art. 49 and deletes only Annex VIII Section B points 7 and 9.

Operative dates as amended:
- **2 Feb 2025** - Prohibited practices (Art. 5) + AI literacy (Art. 4), in effect
- **2 Aug 2025** - GPAI model obligations, in effect
- **2 Aug 2026** - Art. 50 transparency, and Commission GPAI enforcement powers and fines. **Not deferred**
- **2 Dec 2026** - The two new Art. 5 prohibitions (non-consensual intimate material, CSAM), and the Art. 111(4) deadline for synthetic-content systems already on the market to meet Art. 50(2)
- **2 Aug 2027** - Legacy GPAI models (Art. 111(3)), unchanged
- **2 Dec 2027** - High-risk Annex III obligations. Deferred from 2 Aug 2026
- **2 Aug 2028** - High-risk Annex I regulated products. Deferred from 2 Aug 2027

## Development

```bash
npm install
npm run build        # typescript -> dist/
node test.mjs         # full suite incl. ten agent journeys
node test-claims.mjs  # 124-check matrix: pinned law corpus vs served facts
node test-schemas.mjs # post-serialization output-schema gate
node law/fetch.mjs verify # verify all 4 pinned legal documents
npm --prefix compiler test # deterministic compiler: 6 tests
node evals/front-door/score.mjs evals/front-door/agent-args-1.6.0.json # the front door, measured
npm run dev          # stdio dev server
npm run dev:http     # HTTP dev server
```

## Privacy

The hosted endpoint `https://mcp.lexbeam.com/mcp` runs statelessly: it receives the tool arguments your
assistant sends and the request data (IP address, time, URL, user agent), stores no tool arguments or
results, sets no cookies and runs no analytics. It is hosted by Railway. Run the server locally from npm
and nothing reaches Lexbeam. Full policy:
[PRIVACY.md](https://github.com/lexbeam-software/eu-ai-act-mcp/blob/main/PRIVACY.md).

## Disclaimer

This MCP server is a structured information tool that returns references to and summaries of Regulation (EU) 2024/1689. It provides general information only and does not constitute legal advice. Whether a service falls within a national regulated legal-services regime must be verified against current official sources and, where necessary, with qualified local counsel. This tool cannot replace that review, and its use does not establish a lawyer-client relationship. For implementation support, visit [lexbeam.com/kontakt](https://lexbeam.com/kontakt).

## License

MIT. See [LICENSE](LICENSE). Regulation text summarised in `src/knowledge/articles.ts` and `src/knowledge/annex-iv.ts` is derived from Regulation (EU) 2024/1689 as amended; EUR-Lex content is reused under the conditions of Commission Decision 2011/833/EU (preserve attribution, do not distort the meaning of the source).

## About Lexbeam

[Lexbeam Software](https://lexbeam.com) builds agentic AI for compliance, legal operations, internal audit, and risk workflows.

*Give us one ugly, regulation-heavy workflow. We'll turn it into a working AI system fast.*