eu-ai-act-classifier

Find out which EU AI Act risk tier an AI system falls into, with every obligation cited to its article. Rules-based, offline, and built for a lawyer to confirm; see who it's for.

Status: beta, on PyPI. It encodes Regulation (EU) 2024/1689 and, in part, the amending Regulation (EU) 2026/1744, including its application dates and new Art. 5 prohibitions. Source versions are in the methodology, and legal claims, including what is encoded, in the regulatory claim ledger.

Not legal advice. Output is a first pass for supervised legal review, and a qualified human must review every result before reliance. Example data is synthetic. The engine applies rules to characterised facts and does not decide disputed intended purpose, operator role or factual deployment scope (honest limits).

No install needed: the browser playground runs the same engine, and nothing you paste leaves the page.

Install

Python 3.13+, from PyPI:

pip install eu-ai-act-classifier

With uv, prefix eu-ai-act-classify with uvx --from eu-ai-act-classifier to run it without installing.

Quick start

Classify a profile from stdin, offline and deterministically (more profiles in examples/):

echo '{"name":"CitizenScore","prohibited_practices":["5.1.c"]}' | eu-ai-act-classify -

Add --json for JSON, or --strict to exit 1 in CI when the system is prohibited or needs review.

Or from Python:

from eu_ai_act_classifier import SystemProfile, classify, render_report

report = classify(SystemProfile(name="CitizenScore", prohibited_practices=["5.1.c"]))
print(render_report(report))

More: the full quick start and example output.

MCP server

For Claude Code or Claude Desktop, use eu-reg-mcp: uvx --python 3.13 --from 'eu-reg-mcp[lint,classify]' eu-reg-mcp. This package's own MCP server starts with:

uvx --python 3.13 --from "eu-ai-act-classifier[mcp]" --with "mcp<2" eu-ai-act-classifier-mcp

Keep mcp<2 (why). Client configuration: docs/mcp-configuration-guide.md.

Documentation

• Docs index: every guide
• Overview: audience, playground, honest limits, human-authored legal judgment and related tools
• What it checks: scope, Art. 5 prohibited practices, Art. 6 and Annex III high-risk, GPAI and transparency, with Annex III coverage
• Surfaces: playground, CLI, MCP server, local API and web cockpit
• JSON Schemas for profiles and reports
• Launch readiness: a quick evaluator route

License

MIT, see LICENSE.