One deterministic command measures your AI coding harness — works with Cursor, Claude Code, Windsurf, and other tools. Scans AGENTS.md, rules, hooks, tests, CI. No AI in the scan. No network. Same score every time.
The problem
Two repos can run the same model and get wildly different results. One has guides that steer the agent, sensors that verify its work, and guardrails that stop damage — the other has none. Harness Score measures that harness in seconds, across any AI tool, and tells you what to fix next.
How it works
Run npx harness-score in your repo. The CLI reads your filesystem — 36 checks, zero LLM calls, zero network.
Get a maturity level L0–L4, a 108-point breakdown across six dimensions, and the exact gap blocking the next level.
Every failed check links to a remediation recipe in the guide — or let an editor plugin's /harness-audit command apply the fixes.
Example output
Deterministic: same commit, same score — on your laptop or in CI. Gate merges with --min-level 3 so maturity only ratchets up.
Run the scanner → harness-score v1.0.0 ~/my-app
Maturity: L2 · Guided Score: 70/108 (65%)
Detected: Cursor, Claude Code
Context & Guides ████████████████░░░░ 80%
Skills & Commands █████████████░░░░░░░ 65%
Hooks & Guardrails ░░░░░░░░░░░░░░░░░░░░ 0%
Sensors & Feedback ████████████████░░░░ 80%
CI Feedback ██████████████░░░░░░ 71%
Hygiene & Safety ███████████████░░░░░ 74%
To reach L3: sensors ≥ 60%; ci ≥ 50%Get started
Maturity ladder
Levels gate on the shape of your harness — not just points. Eighty points of docs with zero tests is L1, not L3.
L0 · UnharnessedCold-start every session
L1 · DocumentedAGENTS.md orients the agent
L2 · GuidedRules, skills, hygiene
L3 · SensingTests, types, CI verify work
L4 · Self-correctingHooks close the loopWorks with
One harness, one maturity model, many tools. Configure Cursor, Claude Code, Windsurf, or any other AI tool — Harness Score measures them all the same way.
Show your score
112×20 pill for shield rows. CI regenerates it, or pin a static badge-lN.svg. Copy-paste embeds in Markdown, HTML, iframe, and JSX.
What ships in this repo
8 chapters on harness engineering for AI coding agents — feedforward, sensors, guardrails.
harness-score — JSON, markdown, badge output. Zero runtime deps.
/harness-audit command + skill to fix every gap the scan finds.
Scan on every push, emit the badge, fail below --min-level.