Executive summary
What DocSutra AI is, the problem it solves, what you get, and what it deliberately does not claim.
The problem
Documentation rots silently. Files move, headings are renamed, flags are removed, and nobody notices until a reader hits a 404 or copies a command that no longer exists. Manual review does not scale, and most link checkers ignore anchors, case-sensitivity, accessibility and leaked secrets.
The product
DocSutra AI is a command-line auditor for Markdown documentation. One command scans a repository and produces an audit report as Markdown, HTML, JSON and SARIF.
- Offline and private. No network calls, no API keys, no telemetry. The scan never leaves the machine.
- Safe by design. It never executes a command found in a document and never fetches a URL. Credentials it finds are redacted in every output.
- Transparent. Each report includes a verification ledger (what was checked, how many passed and failed) and an explicit "not checked" section. It admits what it did not verify.
- CI-ready. Exit codes 0/1/2, a
--fail-onthreshold, baseline comparison to fail only on new problems, and SARIF for GitHub code scanning. - Agent-ready. 16 portable skills for Claude Code, Cursor, Codex, Gemini CLI and GitHub Copilot, each labelled with what actually backs it.
Who it is for
| Audience | Value |
|---|---|
| Open-source maintainers | Catch broken links and stale anchors before contributors do |
| Technical writers | A readable HTML report with line-level evidence and fixes |
| CI / release engineers | Gate pull requests on newly introduced documentation errors |
| Security reviewers | Find leaked tokens and risky install snippets before publishing |
| AI-assisted teams | Give coding assistants a deterministic scanner plus guardrailed skills |
Evidence at a glance
72tests passed
0tests failed
3skipped (OS privilege)
100%benchmark precision
100%benchmark recall
18rules
Generated 2026-11-16T17:55:30.611Z on Node v24.15.0, win32/x64.
All numbers come from real runs (npm run build:evidence). See Testing and evidence for method and limits.
What it does not claim
- It does not prove documentation is true. A pass means the listed static checks passed.
- It is not a security audit, certification, or WCAG conformance statement.
- Benchmark accuracy is measured on a small author-built corpus, not on arbitrary repositories.
- Skill activation inside each IDE is not verified; files are copied to documented folders only.
- It does not check external URLs, run snippets, or analyse MDX, AsciiDoc or notebooks.
Decision summary
| Question | Answer |
|---|---|
| License | Apache-2.0, zero runtime dependencies |
| Runtime | Node.js 20 or newer |
| Cost | Free; no external services |
| Data leaves machine? | No |
| Maturity | Version 0.x, active; see roadmap |
| Creator | Vatsal Shah |