mcpField
L
Legalithm

Legalithm

Local
github.com/legalithmLegal & Government1Updated 2w ago

About Legalithm

Legalithm is an EU AI Act compliance tool for developers building AI features in an editor. Its offline MCP server classifies risk, explains obligations, and drafts disclosures; it can also create cited compliance records and check them for changes.

Key features

  1. 01Offline risk classification and explanation
  2. 02Cited regulatory obligations
  3. 03Article 50 disclosure drafting
  4. 04Marking and verifying AI content
  5. 05Creating and checking compliance records

Use cases

  1. 01Assess an AI feature's EU AI Act risk tier
  2. 02Mark AI-generated content
  3. 03Detect compliance record drift in CI

MCP support

It provides an offline MCP server with classify, explain_obligation, generate_disclosure, and check_record tools. The first three run fully offline; check_record reads a public API.

Configuration

Local server config is the same for every client. Copy it into your client’s MCP config file.

mcp.json · Universal config
{
"mcpServers": {
"legalithm": {
"command": "npx",
"args": [
"-y",
"legalithm-mcp-server"
]
}
}
}
Run commandnpx -y legalithm-mcp-server

Local servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.

Packages
npmlegalithm-mcp-server0.1.2

Tools

4

Saved tool list · updated 2w ago

ParameterTypeRequiredDescription
audienceenumYesgeneral | workers | children | vulnerable-groups | other
domainenumYesbiometrics | employment | credit | medical | education | law-enforcement | other
roleenumYesprovider | deployer
use_casestringYes

Details

From the project README

Legalithm — EU AI Act compliance in your coding loop

npm version Add to Cursor

Shipping an AI feature to EU users? Article 50 transparency duties have applied since 2 August 2026. Content marking for systems placed before that date is due 2 December 2026, and Annex III high-risk obligations follow on 2 December 2027. Catch it where you code, in seconds.

Source code on the left, connected by four branching lines to a sealed legal document on the right.

This image is AI-generated, and it is marked as such with our own tool: legalithm mark --watermark added a C2PA content credential and a pixel watermark, which is what Article 50(2) asks for. Download it and run legalithm verify on it. The credential is signed with the CLI's test certificate, so it is valid but not trust-listed.

Quickstart

Install the offline server in your editor. No API key, and nothing leaves your machine.

claude plugin marketplace add legalithm-org/legalithm
claude plugin install legalithm@legalithm

Codex:

codex plugin marketplace add legalithm-org/legalithm
codex plugin add legalithm@legalithm

Cursor: use the Add to Cursor badge above.

Then ask your agent "does the EU AI Act apply to this feature, and what tier?"

What comes back (real output, CV screening as the provider)
{
  "risk": "high",
  "confidence": "high",
  "rationale": "This AI system is classified as high-risk under Article 6(2) as it falls into the category of Employment, Workers Management and Access to Self-employment as specified in Annex III...",
  "citations": [
    {
      "article": "6(2)",
      "annex": "III",
      "label": "Article 6(2) & Annex III - High-Risk AI Systems",
      "url": "https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=OJ:L_202401689#article-6",
      "asOf": "2026-08-03",
      "appliesFrom": "2027-12-02"
    }
  ],
  "matchedRules": ["high_risk_employment", "high_risk_domain_selected"],
  "applicableDeadline": "2027-12-02",
  "obligationsHint": {
    "count": 10,
    "topTitles": [
      "Quality Management System (Article 17)",
      "Data Governance (Article 10)",
      "Technical Documentation (Article 11)",
      "Human Oversight (Article 14)"
    ]
  },
  "confidenceScore": 0.72,
  "reviewRequired": false
}

Note matchedRules, asOf and confidenceScore. You can see which rule fired, how old the corpus is, and how sure the engine was. Below the abstention threshold it sets reviewRequired and tells you to get a human instead of guessing.

Content marking for Article 50(2), also no key:

npm i -g legalithm
legalithm mark ./out.png --watermark   # writes out.signed.png
legalithm verify ./out.signed.png      # detect both layers

Install it rather than using npx for these two: C2PA and the watermark come from c2pa-node and sharp, which are optional native dependencies that npx does not reliably fetch. Without them mark warns and marks nothing. Everything else in this README works fine under npx.

The compliance record (needs a free key)
npx legalithm setup   # wires hooks, editor rule and MCP config
npx legalithm init    # writes a dated, cited compliance/legalithm.json
npx legalithm check   # re-verify; non-zero exit on drift (for CI)

init and check talk to the hosted record service, so they need a free API key. Everything above this line does not.

Three surfaces

  1. Editor — an offline MCP server (legalithm-mcp-server) exposing 4 tools (classify, explain_obligation, generate_disclosure, check_record). No API key. The first three run fully offline; check_record reads a public API.
  2. Repo — legalithm init writes a dated, cited compliance/legalithm.json that records your AI system's risk tier and the obligations behind it.
  3. CI — legalithm check and the GitHub Action fail the build when the committed record drifts — because your app changed or the law changed under you.

MCP config

Add the offline server to Claude Code, Cursor or Codex manually:

{
  "mcpServers": {
    "legalithm": {
      "command": "npx",
      "args": ["-y", "legalithm-mcp-server"]
    }
  }
}

GitHub Action

# .github/workflows/ai-act.yml
name: AI Act
on: [pull_request]
jobs:
  ai-act:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: legalithm-org/legalithm/packages/action@v1
        with:
          api-key: ${{ secrets.LEGALITHM_API_KEY }}

Honest framing

A cited starting point that tells you when to get a human — not legal advice. When unsure, it flags the result for review instead of guessing. Every output is checked against Regulation (EU) 2024/1689; it is not a certification.

Links

License

MIT

Comments

Markdown supported

New to MCP?

From what the protocol is to installing your first server, these answers cover most first questions.

An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.