MCP adapter

judge-mcp lives in the judge-mcp/ directory of the repository. It is a thin stdio adapter: every tool call runs an already-built cargo-judge binary and returns its --format json output to the MCP client. No state, no cloud, no network beyond that local subprocess. It is not published to a registry, and it is not a requirement of judge.

In practice

1. Health check: “How healthy is this workspace before the release?” health with score: true reports a score of 90.0 (grade A) for 20,493 lines, no failing findings and 68 warnings, with 72 gating and 54 advisory findings. The most frequent rules are swallowed-result (33), panic-in-lib (32) and silent-default (20). The lowest maintainability index (1.0) is in origin-xtask/src/validate.rs.
2. Duplicates: “Where do we have duplicated code?” dupes reports 62 clone families with 126 members. The largest is SyncEngine::sync_if_due (engine.rs:253) and SyncEngine::sync_now (engine.rs:286), 108 tokens each. One family crosses crates: proxy_streams (origin-mcp-http/src/proxy.rs:39) and serve_streams (origin-mcp-stdio/src/lib.rs:37), 63 tokens each.
3. Delete or keep? “Can I delete demo_refresh in the demo app, or is it still used?” dead_code reports it as unused-pub-workspace with 0 references, and lists the limitation proc_macro_expansion_disabled. inspect_symbol finds no path to demo_refresh (reachable: false). For origin-demo::run it shows a static call from main (main.rs:5). Because run registers the command inside a macro, which judge does not expand, the assistant keeps the function.
4. Explain a finding: “What does the silent-default finding in origin-auth/src/token.rs:91 mean?” explain_finding describes the rule as a heuristic with the verdict effect advisory_only, to be stated “as a hint or possible reading, never as proof”. It also gives its exclusions: a syntax-only proxy, not a taint proof.
5. Refactoring plan: “cargo judge patterns suggests pattern:raii-guard:51769d26d9a7fc5e. How would I refactor that?” fix_preview returns a four-step migration plan for raii_guard. Step 1 is a manual check, and step 4 points to crates/origin-platform/src/shortcut.rs. The plan comes with patch: null, because judge generates no patch.

Build

cd judge-mcp
npm install
npm run build   # produces dist/index.js

It needs Node.js 18 or newer and the cargo-judge binary. JUDGE_BINARY sets the path to the binary; by default cargo-judge must be on PATH. If the binary cannot be started, the affected tool call returns an error explaining how to fix it; the server itself still starts.

Client configuration

{
  "mcpServers": {
    "judge": {
      "command": "node",
      "args": ["/absolute/path/to/judge-mcp/dist/index.js"],
      "env": {
        "JUDGE_BINARY": "/absolute/path/to/cargo-judge"
      }
    }
  }
}

Tools

Tool	Runs
analyze	cargo-judge --format json
health	cargo-judge health --format json [--score]
dupes	cargo-judge dupes --format json [--mode ...]
dead_code	cargo-judge dead-code --format json (needs a --features deep build)
explain_finding	cargo-judge explain-rule <rule-id> --format json
inspect_symbol	cargo-judge explain <item-path> --why-live --format json (needs a --features deep build)
fix_preview	cargo-judge fix-preview <pattern-id> --format json

All tools are read-only; cargo-judge never modifies code. explain_finding explains the rule behind a finding (evidence class, preconditions, exclusions, verdict effect), not the individual occurrence, because the CLI has no per-finding explain command.