Clarx measures how safely and efficiently an AI agent can navigate, modify, and verify changes in your repository — five pillars, 27 rules, and a score that says what it is, what it is not, and what it could not evaluate.
The number answers one question: can an agent find its way, stay inside boundaries, and verify its own changes here? It never pretends to audit quality or security.
Rules that cannot run on your stack are reported as not evaluated, never silently passed. A Python repo sees exactly which import-graph checks did not apply — and they never move the score.
All numeric limits ship with written reasoning and can be overridden per repo in clarx-manifest.json. When you tune one, reports render the active value — tuned repos read honestly.
Blocking issues impose a graduated score floor and a fix-first list that names the cap each one carries — so the path from 65 back to 90 is explicit, not mysterious.
Five pillars — discoverability, boundary clarity, context efficiency, operational guidance, edit safety — measured by 27 rules. Public, versioned, and language-agnostic.
OPEN SPECIFICATIONnpx @clarxai/cli score . runs locally or in CI, explains every rule on demand, and reads your manifest for overrides. No account required.
Scan history, org-wide benchmarks, AI-enriched findings review, and a manifest studio with live pillar scoring — the team workflow on top of the same engine.
HOSTED WORKFLOWThe real analyzer, running in your browser. Nothing leaves this page.
Open the studio →One command, no config. The report leads with what to fix first, names the cap each hard failure imposes, and declares which rules could not be evaluated for your stack — inapplicable rules never move the score.
The score is called what it is, says what it is not, declares what it could not evaluate, and defends its thresholds.
Install the CLI, run your first scan, connect to CI.