default
Deterministic Scanner (<2s)
Inspects files, scripts, hooks, CI workflows, PR templates and git history locally. No LLM hallucination, repeatable in CI.
OPEN-SOURCE CLI · MIT · 100% LOCAL
Run one command inside any repository to inspect your real agentic setup — context files, hooks, sandboxing, worktree hygiene, test oracles and CI gates. Zero API keys, zero code uploaded, deterministic Level 0–5 score.
Start with the zero-config deterministic scan, then add qualitative review (--deep), a 7-question team interview (--interview), or a tailored fix prompt (--prompt).
Deterministic local audit (<2s, zero API keys, zero network calls)
npx @wondelai/harness-audit --format html --out report.html for a shareable standalone report.--prompt to emit HARNESS-FIX-PROMPT.md tailored to your exact missing checks.npx @wondelai/harness-audit init for a minimal, non-destructive starter scaffold.Deep audit via your local Claude Code CLI + companion skill
npx @wondelai/harness-audit --deep --prompt claude CLI — no separate API key needed.CLAUDE.md size/bloat, .claude/settings.json permissions, and PreToolUse/Stop hooks./harness-audit skill in wondelai/skills.Checks AGENTS.md hierarchy, sandbox_mode and approval_policy
npx @wondelai/harness-audit --deep --provider codex AGENTS.md placement, size bounds, and explicit build/test verification commands..codex/config.toml for safe sandbox_mode and approval_policy separation.--approval-mode or untrusted policy leftovers.Checks .cursor/rules/*.mdc scoping and Bugbot review config
npx @wondelai/harness-audit --format md --out AUDIT.md .cursorrules files and validates scoped .cursor/rules/*.mdc frontmatter..cursor/BUGBOT.md automated review criteria on pull requests.CLAUDE.md pointing via @AGENTS.md so all three tools share one truth.Fail the build if harness hygiene regresses below your threshold
npx @wondelai/harness-audit --ci --min-score 70 --format json --ci and --format json modes.Everything the old web scorecards and static Setup Pack tried to do, rebuilt directly against your real repository.
default
Inspects files, scripts, hooks, CI workflows, PR templates and git history locally. No LLM hallucination, repeatable in CI.
--deep
Optionally invokes your own local claude/codex CLI or API key to critique rule bloat, vague acceptance criteria and brittle tests.
--interview --role dev|lead|cto
Adds the human and organizational dimensions files cannot show: review queues, risk-class merge policies and outcome metrics.
--prompt / init
Generates HARNESS-FIX-PROMPT.md for your coding agent, or runs init to scaffold minimal AGENTS.md, CLAUDE.md and PR template without overwriting.
Explore a realistic 6-station audit below — or paste/drop your own report.json from npx @wondelai/harness-audit --format json. Parsing runs 100% inside your browser.
acme/billing-api (sample)
Sample repository auditHarness Score
64/100
Claude Code · Cursor
Autonomy Level
L3
Review Manager
Capping Station
Verification
Active Contributors
8
PR template and architecture notes exist; missing explicit risk-class labels.
AGENTS.md & CLAUDE.md present, but Stop/PreToolUse hooks are not configured.
Single-command test and typecheck scripts verified in package.json.
CI runs tests, but worktree isolation and port hygiene are undocumented.
No oracle protection rule in CI and no architectural fitness functions.
CI pipeline runs on PRs, but no automated AI PR review or evidence bundle.
Protected tests can be weakened by agents without failing CI
Add a CI guard or PreToolUse hook that flags edits to contract/oracle tests.
Read free chapter → Protect the OracleNo deterministic enforcement hooks configured (PreToolUse / Stop)
Move critical formatting, lint and secret checks from prompt prose into hooks.
Read free chapter → Agent Hooks as GuardrailsPull requests lack a structured evidence bundle checklist
Require test exit status, verification commands and risk class in .github/pull_request_template.md.
CLAUDE.md duplicates AGENTS.md instead of referencing @AGENTS.md
Keep a single source of truth in AGENTS.md and import it with @AGENTS.md.
100% client-side JSON.parse — nothing leaves your browser.
Run npx @wondelai/harness-audit --format json > report.json and paste or load the file here. Zero data is sent to any server.
FOR CTOS, VPS OF ENGINEERING & CEOS
You do not need to clone a repo yourself to see where your engineering organization actually sits on the Autonomy Ladder.
Ready-to-paste message (100% local, zero source code leaves the company):
Hey — could you run `npx @wondelai/harness-audit --format html --out harness-report.html` on our main repo and send me the HTML summary? It runs 100% locally (MIT, zero code upload) and gives us a clean baseline of our AI harness & Level 0–5 bottlenecks.
Michał Jaskólski audits your main repository directly (written findings + prioritized 30-day roadmap from $499) or spends a full Working Day implementing hooks, CI gates and review standards with your team ($2,499).
No. The default scan is 100% local, deterministic, offline-capable and sends zero telemetry or source code anywhere. The web JSON viewer on this page also parses everything locally in your browser with JSON.parse.
--deep mode work without vendor lock-in? --deep automatically detects a locally installed claude or codex CLI (using your existing session) or accepts your own ANTHROPIC_API_KEY / OPENAI_API_KEY. It only runs when you explicitly pass --deep.
A 25-question web survey measures self-perception; harness-audit inspects your actual repository and adds an optional 7-question --interview for org dimensions. And --prompt + init replace the old static Setup Pack ZIP with a tailored fix prompt based on what your repo actually lacks.
Yes — MIT licensed, free for solo developers and enterprises alike. If you want offline books for your team or hands-on implementation help, those are available on /en/books/ and /en/services/.
One tap, no sign-in.
What are you here for?
What is holding you back?
How does it read?
Thanks. If you left an email, Michał will reply.
Answering your questions?
One tap each. Close whenever you like.
no sign-in
Run npx @wondelai/harness-audit in two seconds, read the free docs chapters for your top gaps, or bring us in for a fixed-price Repo Diagnostic.
Michał replies, not a bot. The address is used for this reply only.