license-sentinel

license-sentinel

Audit Python and npm dependency licenses for compliance before you ship — an MCP server for AI coding agents.

An AI agent can add pdf-renderer to your project in one second. It will not tell you that pdf-renderer is AGPL-3.0 and that shipping it inside a closed-source product is a license violation. License data and compatibility rules are things a model cannot reliably recall — packages relicense between versions (MongoDB → SSPL, Redis → BUSL, Elasticsearch → Elastic-2.0), and "the source is on GitHub" does not mean "free to ship".

license-sentinel reads what is actually on your disk and judges it against how you distribute your product.

• Works for Python and npm in one pass — existing MCP license tools are npm-only.
• Runs locally over stdio. No network calls, no telemetry, nothing leaves your machine.
• Verdicts, not raw data: CLEAN / REVIEW / BLOCK, each with the reason in plain language.

Tools

Tool	What it does
audit_project(path, context)	Scan a project's dependencies and return counts plus every BLOCKING and REVIEW item with reasons.
check_package(names, context)	Check specific packages or messy license strings before installing. Accepts AGPL-3.0, BUSL-1.1, GPLv3, Apache License 2.0, MIT OR Apache-2.0.
generate_notices(path, output)	Write a THIRD-PARTY-NOTICES.md attribution document for client hand-off.

There is also a pre_release_license_review prompt that chains the audit into a go/no-go review.

Install

# run without installing (recommended)
uvx --from license-sentinel license-sentinel

# or install
uv pip install license-sentinel
# or
pip install license-sentinel

Configure your client

Claude Desktop / Cursor / Windsurf / VS Code Copilot / Zed all read the same shape:

{
  "mcpServers": {
    "license-sentinel": {
      "command": "uvx",
      "args": ["--from", "license-sentinel", "license-sentinel"]
    }
  }
}

If you installed with pip instead, use "command": "license-sentinel" with no args. Restart the client and the three tools appear.

Distribution context

The same dependency is fine in one context and fatal in another, so every tool takes a context argument:

Context	Meaning	What it blocks
proprietary (default)	Closed-source product you distribute	GPL/AGPL/SSPL, BUSL/Elastic, non-commercial
saas-backend	Never distributed, only runs on your servers	AGPL/SSPL (network trigger), BUSL/Elastic
permissive	Your own project is MIT/Apache/BSD	Anything copyleft that would contaminate your terms
copyleft-ok	Your own project is GPL family	Only source-available and non-commercial

What it reads

• Python: .venv/ / venv/ / env/ installed packages (dist-info/METADATA), requirements.txt, pyproject.toml (PEP 621, poetry, dependency-groups)
• npm: node_modules/*/package.json (including scoped packages), package.json dependencies

If a dependency is declared but not installed, it is reported with an UNKNOWN license rather than silently dropped — an unlicensed dependency is all-rights-reserved by default.

Privacy

No HTTP client is imported anywhere in this package. The scan is read-only (except generate_notices, which writes the file you name). Nothing is uploaded.

Limitations

• Not legal advice. It is a fast first pass that catches the expensive mistakes; have counsel review anything flagged.
• Transitive dependencies are read from what is installed. If you have no .venv and no node_modules, declared-only dependencies come back UNKNOWN.
• The current environment running the server is never scanned, so the server's own packages never pollute your report. Set LICENSE_SENTINEL_SCAN_CURRENT_ENV=1 to change that.

Development

uv sync
python tests/smoke_test.py     # 9 tests, no pytest needed
python tests/e2e_check.py      # calls the tools end to end

License

MIT