45 specialized judges evaluating AI-generated code for security, cost, and quality
npx @kevinrabun/judges-cli eval --file src/app.ts JUDGES_SHOW_EXPERIMENTAL=1 judges --help
uses: KevinRabun/judges@main with: fail-on-findings: true upload-sarif: true
npm i @kevinrabun/judges
import { evaluateCode } from "@kevinrabun/judges/api";
judges baseline create --dir . judges license-scan --dir .
Real-world analysis of popular open-source projects using the deterministic tool layer (regex + AST, zero LLM calls).
| Repository | Verdict | Score | Files | Critical | High | Total |
|---|
| Step | Description |
|---|---|
| 1. Parse | Source code is parsed with regex patterns and TypeScript AST analysis — instant and offline |
| 2. Evaluate | All 45 judges independently score the code on their domain (0-100) using deterministic heuristics |
| 3. Tribunal | Scores are aggregated into a weighted verdict: PASS / WARNING / FAIL |
| 4. Report | Findings include evidence, line numbers, confidence, and auto-fix patches |
| 5. Deep Review | 45 expert-persona MCP prompts let your LLM client perform probabilistic analysis beyond static patterns |
| Method | Command |
|---|---|
| CLI | npx @kevinrabun/judges eval --file app.ts |
| MCP Server | npm install @kevinrabun/judges — add to your MCP client config |
| GitHub Action | uses: KevinRabun/judges@main — see README |
| Docker | docker run --rm -v $(pwd):/code judges eval --file /code/app.ts |
@kevinrabun/judges-cli → provides the judges CLI binary.@kevinrabun/judges → MCP server + programmatic API (evaluateCode, evaluateCodeV2, etc.).npx @kevinrabun/judges-cli ....
Audit dependency licenses and flag copyleft / unknown licenses.
judges license-scan --format jsonjudges license-scan --risk highjudges license-scan --category weak-copyleft --save.judges-licenses/license-report.json when --save is used.