sovereign-ai-act-mcp

<div align="center">

# ◆ Sovereign AI Act — MCP server

**Give your AI agent a *trustworthy* EU AI Act expert.**

[![npm version](https://img.shields.io/npm/v/sovereign-ai-act-mcp?color=4DD6E0&label=npm)](https://www.npmjs.com/package/sovereign-ai-act-mcp)
[![npm downloads](https://img.shields.io/npm/dm/sovereign-ai-act-mcp?color=4DD6E0)](https://www.npmjs.com/package/sovereign-ai-act-mcp)
[![license](https://img.shields.io/npm/l/sovereign-ai-act-mcp?color=C9A961)](./LICENSE)
[![MCP](https://img.shields.io/badge/Model_Context_Protocol-compatible-4DD6E0)](https://modelcontextprotocol.io)

</div>

This [Model Context Protocol](https://modelcontextprotocol.io) server exposes **Leo** — the deterministic classifier for **Regulation (EU) 2024/1689 (the EU AI Act), as amended by Regulation (EU) 2026/1744 (the Digital Omnibus on AI, in force 27 July 2026)** — as callable tools. Every answer is grounded **verbatim** in the official law. **Leo never guesses.**

When a user asks an AI *"is my AI system high-risk under the EU AI Act?"*, the honest answer must cite the actual law — not a hallucination. This server lets your agent do exactly that.

## 🛠️ Tools

| Tool | What it does |
|------|--------------|
| `classify_ai_system` | Plain-language AI description → risk tier (prohibited / high-risk / limited / minimal) + the exact Annex III category and binding Articles. |
| `lookup_article` | Verbatim direct lookup for original Articles 1–113 plus inserted Articles 4a, 60a and 75a–75d. |
| `search_eu_ai_act` | Full-text search across the original 2024 Articles, Recitals and Annexes. |
| `get_compliance_deadlines` | The application dates in force under Regulation (EU) 2026/1744 (statuses computed at call time) and the Article 99 fine tiers. |
| `review_feature` | Pre-release review of an AI feature: signal, binding Articles, obligations, red flags and next steps. |

## 🚀 Install (Claude Desktop)

Add to your `claude_desktop_config.json`:

```json
{
  "mcpServers": {
    "sovereign-ai-act": {
      "command": "npx",
      "args": ["-y", "sovereign-ai-act-mcp"]
    }
  }
}
```

Restart Claude, then ask: *"Is an AI that screens job applicants high-risk under the EU AI Act?"* — Claude calls Leo and answers with the exact Articles.

Works the same in **Cursor**, **VS Code**, **Windsurf**, **Cline**, and any MCP-capable client.

## 🌐 Remote server (zero install)

Clients that support Streamable HTTP can use the same five tools directly:

```text
https://www.regulatoryai.eu/mcp
```

## ⚡ Try the API directly (no install, no key)

```bash
curl -X POST https://www.regulatoryai.eu/api/classify \
  -H "content-type: application/json" \
  -d '{"description":"AI that screens job applicants and ranks CVs"}'
```

```json
{
  "ok": true,
  "tier": "high_risk",
  "classifications": [{
    "tier": "high_risk",
    "rule": "Employment, recruitment & worker management (Annex III.4)",
    "annex": "III.4",
    "articles": [6, 9, 10, 11, 13, 14, 15],
    "severity": "HIGH — conformity assessment + risk management + ..."
  }]
}
```

Also: `GET /api/article/{1-113|4a|60a|75a|75b|75c|75d}` · `GET /api/annex/XIV/1` · `GET /api/search?q=...` (original 2024 corpus plus the listed 2026 insertions)

## 💎 Why trust it

- **Verbatim law.** The original 2024 corpus contains 113 Articles, 180 Recitals and 13 Annexes. Inserted Articles 4a, 60a, 75a–75d and Annex XIV from Regulation (EU) 2026/1744 are published separately in all 24 languages.
- **Deterministic.** A rule engine, not a guess — every verdict cites the Article that binds you.
- **EU-sovereign.** Operated by Dominion Intelligence AB, hosted in the EU.
- **24 languages.** Ask in your own language.

## ⚙️ Config

- `SOVEREIGN_API_BASE` — override the API base URL (default `https://www.regulatoryai.eu`).
- No API key required.

---

Indicative classification grounded in the official text — **not legal advice**.
Powered by [**Sovereign AI Act**](https://www.regulatoryai.eu/for-ai/) · © Dominion Intelligence AB