shipcheck-mcp

Find what your AI-built app leaks — before you ship.

Works with Claude Code, OpenCode, Cursor, or any MCP-compatible agent — it's a standard Model Context Protocol server, not a Claude-only tool.

AI coding agents happily ship apps that leak API keys, pull in GPL-licensed packages, leave admin routes open to the internet, collect emails with no privacy policy, or touch raw card numbers — because "does it run" and "is this okay to ship" are different questions, and your agent only answers the first one.

ShipCheck answers the second one. It runs five fast, deterministic checks over your project and explains every finding in plain words — what's wrong, why it matters for you, and exactly how to fix it. No security background needed.

Honest by design: ShipCheck flags risky patterns. It is not a legal review and not a guarantee. A clean scan means nothing matched these specific checks — nothing more. See PRD.md §2 for the full non-goals.

Two tools, two moments

The bundled skill teaches your agent the habit: scan_diff before commits, scan_repo before launch.

Related MCP server: agentguard

Works with your stack

Every summary tells you which stacks it detected (project_types) and warns you honestly when a detected stack isn't fully covered (coverage_caveat) instead of implying everything is fine.

What a finding looks like

[critical] lib/firebase_options.dart — Possible Google API Key found in code that runs on people's devices.

Why this matters: This key travels inside your app to every user's phone or browser. Anyone can copy it out and use it as if they were you — spending your money or reading your data.

The fix: Move the key into a separate secrets file that never becomes part of the project, load it at startup, then get a new key from the provider's dashboard — treat this one as stolen.

30-second start

# In Claude Code:
/plugin marketplace add noureldeensalama/shipcheck
/plugin install shipcheck

# Then just talk to your agent:

"run shipcheck before I commit" · "is this ready to launch?" · "scan this repo"

Headless / CI:

claude --plugin-dir /path/to/shipcheck -p "Run shipcheck scan_repo on . and summarize findings"

Standalone MCP server (Cursor, OpenCode, anything else)

git clone https://github.com/noureldeensalama/shipcheck.git
cd shipcheck && npm install && npm run build

{
  "mcpServers": {
    "shipcheck": { "command": "node", "args": ["/absolute/path/to/shipcheck/dist/index.js"] }
  }
}

Or once published: npx shipcheck-mcp

Built to be trusted

• Deterministic: regex + dependency metadata only. An LLM explains findings downstream; it never decides whether a finding exists (PRD.md §4).

• Quiet by design: identical keys deduplicate into one finding with all locations; output is severity-ordered, compact (~1k tokens on a 100k-LOC repo), capped at 100 findings with an explicit truncation flag; vendored trees (node_modules, .venv, build…) are never scanned; placeholder values (mock…, CI dummies) and public-by-design configs (google-services.json, firebase_options.dart) don't cry wolf.

• Proven on real code: dogfooded on five production repos — it caught a live Supabase service-role key across 13 files and three unauthenticated debug endpoints, then verified the fixes landed and flagged a second key that survived only in git history. Full true/false-positive log: DOGFOOD_RESULTS.md.

Compatibility

Testing it yourself

npm test             # 53 unit tests
npm run e2e          # 21 real-MCP-protocol checks
npm run verify-fixture   # acceptance gate over deliberately-vulnerable fixtures
npm run dogfood -- /path/to/any/repo   # scan any repo from the CLI

Contributing

Issues and PRs welcome — see CONTRIBUTING.md. New detector categories need both a fires-on-risk fixture and a doesn't-over-fire regression test; false-positive reports with real code samples are especially valuable. Real-repo results live in DOGFOOD_RESULTS.md.

License

MIT — see LICENSE.