EU AI Act MCP Server

An open-source Model Context Protocol (MCP) server that gives LLMs structured intelligence about the EU AI Act (Regulation (EU) 2024/1689).

Built by Lexbeam Software — an agentic AI implementation boutique for regulated workflows.

What's new in 1.1.5

• Annex III(5) FRIA citation labels corrected. Article 27(1) universal FRIA triggers are now correctly cited as Annex III(5)(b) (creditworthiness or credit scoring) and Annex III(5)(c) (risk assessment and pricing for life and health insurance). Previous labels (5(a)/5(b) or 5(b)/5(a)) are fixed across articles.ts, faq-database.ts, obligations.ts, and the generated dist.

• Article 27 carve-out clarified. Annex III point 2 (critical infrastructure) is the only Annex III category exempt from FRIA, now stated explicitly in every relevant surface.

• Article 43 conformity assessment text corrected. Previous wording suggested "certain critical infrastructure" required notified-body involvement. Fixed: Annex III points 2-8 follow internal-control under Annex VI per Art. 43(2). Notified-body involvement is the Annex III(1) biometrics route under Art. 43(1) and the Annex I sectoral route under Art. 43(3).

• Version skew resolved. package.json, MCP server metadata (src/server.ts), and /health endpoint now all report 1.1.5 consistently.

• Test count. Local suite is now 110 tests passing.

Related MCP server: ai-economy-infrastructure

What's new in 1.4.3

• Corrected a mechanism claim that was wrong as law. The server described 2 December 2027 as a backstop that high-risk obligations could beat by six months after a Commission decision confirming that supporting standards and support measures were available. That conditional trigger comes from Commission proposal COM(2025) 836. It was deleted before adoption. Art. 113, third paragraph, point (c) as replaced by item 40 of Regulation (EU) 2026/1744 sets two plain calendar dates with no condition, so 2 December 2027 and 2 August 2028 are fixed and nothing can pull them forward. The wrong version shipped in 1.4.0 through 1.4.2.

• Renamed the timeline fields so the names cannot re-teach the error. In euaiact_check_deadlines with include_pending_omnibus: true, high_risk_timeline.mechanism is now superseded_proposal_mechanism and high_risk_timeline.backstop is now application_dates, each with its *_source_status renamed to match. The superseded proposal text is retained, clearly labelled, so analyses written from the proposal can be recognised as out of date.

• Nine new guard tests, including a payload-level assertion that no milestone in the default response calls a high-risk date a backstop or offers an earlier support-measures trigger. Suite is 312 tests passing.

• smithery.yaml had drifted to 1.4.1 while package.json was at 1.4.2; both now read 1.4.3.

What's new in 1.4.0

• The Digital Omnibus on AI is enacted. Published in the Official Journal on 24 July 2026 as Regulation (EU) 2026/1744 (CELEX 32026R1744), in force from 27 July 2026. The server now serves the amended dates as operative law.

• Deferred: Annex III high-risk obligations move from 2 August 2026 to 2 December 2027; Annex I product-integrated high-risk obligations from 2 August 2027 to 2 August 2028. Both are unconditional calendar dates. The Commission proposal would have tied application to a decision on the availability of support measures; that trigger was deleted before adoption and is not in Art. 113(3)(c) as enacted.

• Not deferred: Art. 50 transparency and the Commission's GPAI enforcement powers and fines stay on 2 August 2026; the legacy GPAI compliance date in Art. 111(3) stays 2 August 2027.

• New obligations from 2 December 2026: the Art. 5(1)(ba) and (bb) prohibitions on non-consensual intimate material and CSAM, and the Art. 111(4) transition requiring synthetic-content systems already on the market to comply with Art. 50(2).

• Obligation deadlines are now derived, not hardcoded, so euaiact_get_obligations and euaiact_check_deadlines can no longer state different application dates for the same system. A cross-tool consistency test guards this.

What's new in 1.1.4

• Digital Omnibus block in euaiact_check_deadlines updated to reflect the 2026-05-07 Council/Parliament provisional political agreement. status flips from "proposal_only" to "provisional_agreement". The agreement is not yet adopted law — current-law dates remain authoritative for compliance advice until formal adoption plus Official Journal publication.

What's new in 1.1.0

• Structured classifier signals. euaiact_classify_system now accepts optional signals (domain, uses_biometrics, biometric_realtime, is_safety_component_of_regulated_product, generates_synthetic_content, interacts_with_natural_persons, etc.). Signals take precedence over text matching and give deterministic, high-confidence answers on canonical Art. 5 / Annex III / Art. 50 cases.

• Matched signals + follow-up questions. Every classification now returns matched_signals, missing_signals, and next_questions so the calling agent can explain why and ask the user what's still needed.

• Classifier correctness fixes. Rewrote src/utils/matching.ts to eliminate a multi-word-keyword false-positive bug (where a single-character token like "e" in "e-commerce" could match keywords starting with "e") and a fractional-denominator false-negative (where realistic recruitment descriptions scored below threshold on Annex III(4)). See AUDIT.md for root-cause detail.

• 4 new tools:

  • euaiact_get_article — retrieve operational summaries of the most-cited articles (Art. 3, 4, 5, 6, 9-17, 26, 27, 43, 47, 49, 50, 51, 53, 55, 72, 73, 99, 100, 113) plus stable EUR-Lex URLs for grounded citations.

  • euaiact_check_gpai_systemic_risk — determine whether a GPAI model crosses the Art. 51(2) 10²⁵ FLOPs threshold and return Art. 53 baseline + Art. 55 systemic-risk obligations with the Art. 52 notification duty.

  • euaiact_assess_art6_3_exception — walk through the Art. 6(3) "no significant risk" exception with explicit handling of the profiling block (Art. 6(3) second subparagraph) and the Art. 6(4) documentation reminder + Art. 49(2) registration duty.

  • euaiact_annex_iv_checklist — return all nine Annex IV technical-documentation items, optionally as a markdown checklist, with an SME-simplified note.

• 2 new resources: euaiact://annex/iii (full Annex III categories) and euaiact://annex/iv (full Annex IV checklist), joining the existing euaiact://timeline and euaiact://risk-levels.

• New prompt: ground-citation — guides the agent to call euaiact_get_article and quote with an EUR-Lex URL.

• 5 new FAQ entries: FLOPs threshold for systemic-risk GPAI, FRIA for credit scoring, chatbot disclosure under Art. 50(1), minimal-risk spellchecker/recommender examples, and an expanded Art. 6(3) exception entry with the profiling caveat.

• Penalty fix. When is_sme: true the tier_details.description now correctly says "whichever is lower (Art. 99(6) SME/startup protection)" instead of contradicting the max_fine.explanation. A new comparative block shows the SME reduction alongside the non-SME amount.

• Better deadlines tool. New only_upcoming filter and a next_milestone shortcut at the top of the response.

• Improved FAQ search. findBestMatch uses symmetric overlap (matched / min(query_words, item_words)), so specific multi-word queries like "FRIA for credit scoring" no longer drop to generic answers.

• Slim per-response branding. disclaimer, source, and last_updated were moved into the MCP serverInfo.instructions shown once on initialize. Agents no longer pay a per-call context tax for attribution. lexbeam_url is kept only where it adds deep-dive value (FAQ, obligations, classifier).

• Comprehensive test suite. Includes regression tests for every bug fixed in this release. Current local suite count: 110 tests passing.

Tools

Resources

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 (new) — retrieve article text + EUR-Lex URL for grounded citations

Quick Start

npx (no install)

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

Claude Desktop

Add to claude_desktop_config.json:

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

Smithery

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

Direct MCP endpoint: https://eu-ai-act--lexbeam-software.run.tools.

From source

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)

Knowledge Base

Curated, structured data covering:

• 8 Annex III high-risk categories with keyword matching and examples

• 8 prohibited AI practices (Art. 5 (1) (a)-(h))

• Art. 6(3) exception conditions with the profiling block rule

• Art. 50 transparency triggers (chatbots, deepfakes, emotion recognition, machine-readable marking)

• 5 implementation milestones with dynamic days-remaining calculation

• Digital Omnibus proposal status and impact assessment

• Provider obligations (13 for high-risk, 8 for GPAI including Art. 53 + Art. 55)

• Deployer obligations (8 for high-risk)

• Limited-risk transparency obligations (4 under Art. 50)

• Universal AI literacy (Art. 4)

• Penalty framework with SME protection logic (Art. 99)

• 24 FAQ entries with article references and Lexbeam knowledge-base links

• 27 article summaries with EUR-Lex URLs (new in 1.1.0)

• Annex IV (9 documentation items) (new in 1.1.0)

All dates, articles, and obligations verified against the regulation text.

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 item is left unresolved. 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

npm install
npm run build        # typescript -> dist/
node test.mjs        # run the 303-test suite
npm run dev          # stdio dev server
npm run dev:http     # HTTP dev server

Disclaimer

This MCP server is a structured information tool that returns references to and summaries of Regulation (EU) 2024/1689. It is not Rechtsberatung im Sinne des § 2 RDG and does not constitute legal advice in any jurisdiction. It cannot replace consultation with a qualified Rechtsanwalt or equivalent licensed professional. Use of this tool does not establish a lawyer-client relationship. For implementation support, visit lexbeam.com/kontakt.

License

MIT. See LICENSE. Regulation text summarised in src/knowledge/articles.ts and src/knowledge/annex-iv.ts is derived from Regulation (EU) 2024/1689, which is public-domain under Commission Decision 2011/833/EU.

About Lexbeam

Lexbeam Software 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.