dockerfile-audit-mcp

An MCP server that audits Dockerfiles for real container-security anti-patterns. Parses actual Dockerfile structure (instructions, backslash line continuations, multi-stage builds) via a hand-written parser, not regex over the raw text.

What it catches

Missing USER. If the final stage that actually ships has no USER instruction (or explicitly sets USER root), every process in the running container has root privileges by default. Correctly checks only the final stage — matching real linter convention (hadolint's DL3002), since multi-stage builds exist specifically so earlier build-only stages' root steps never ship.

Baked-in secrets. ENV/ARG values assigned to secret-shaped names (API_KEY, PASSWORD, *_TOKEN, STRIPE_*_KEY, etc.) land permanently in the image's layer history — visible via docker history --no-trunc to anyone who pulls the image, even after a later layer unsets the variable. Placeholder-looking values (<your-key>, changeme) and bare ARG declarations with no default are correctly not flagged.

curl | sh / wget | bash. Pipes a remote script directly into a shell at build time with no integrity check — if the host is compromised or the script changes, every future build silently pulls in whatever it now serves.

Unpinned base images. :latest or no tag at all means the base your image builds on can change between builds with nothing in the Dockerfile to explain why.

ADD with a remote URL. Same unverified-fetch problem as curl | sh, via a different instruction.

Tools

audit_dockerfile

Full audit. Returns a risk level and every finding with its exact location, why it matters, and a concrete fix.

Use it

Hosted (recommended): MCPize — free tier, $7/mo Pro.

Self-host:

npm install
node server.js

Part of a small suite

github-actions-audit-mcp, regex-safety-audit-mcp, secrets-leak-audit-mcp, mcp-trust-audit-mcp.

License

MIT