Contributing to mcpgawk

Thanks for helping. mcpgawk is small on purpose. Two rules are non-negotiable — a PR that breaks either will not be merged.

The two invariants

No inventory egress. The measurement layers (measure, label, signals, drift, history) must never import a network library or send anything anywhere. Network lives only in probe.py (the protocol client talking to the scanned server) and servercard.py (public, unauthenticated card fetch — no auth headers, no redirect-following). A test enforces this; don't weaken it.
Zero false positives, or the signal is cut. Any new BOUNDED signal must be precise (it fires on language aimed at the model, never on legitimate capability keywords) and must score 0 false positives on the test corpus before it ships. A capability keyword (url, delete, base64) is a fact (measure.py), never a signal. A signal is a pointer for a human, never a verdict.

Keep the EXACT / INDEX / BOUNDED tiers separate — never let a heuristic contaminate a fact.
Cost is a named tokenizer index, not an absolute count — keep it labelled honestly.
Add a test with every change. Run: uv run --extra dev --with mcp --with tiktoken --with httpx python -m pytest -q.
No telemetry, no account, no "phone home" — ever. That is the whole point of the tool.

git clone <repo> && cd mcpgawk
uv run --with mcp --with tiktoken --with httpx python -m mcpgawk scan examples/mcp.json

By contributing you agree your contribution is licensed under Apache-2.0.