FOR AGENTS AND CI

One command.
Nothing hidden.

A normal CLI, a JSON report, and explicit boundaries. No bespoke agent server to operate.

Start offline

anchorlint audit ./dist   --base-url https://example.com   --format json --output audit.json

Choose only the user-authorized built site. Read the process exit and report status, coverage and errors before summarizing findings.

Keep the evidence intact

Report the exact anchor, surrounding words and destination. Distinguish a missing local page from a live HTTP failure, a model warning from a verified fact, and unevaluated coverage from a pass.

Keep approval explicit

  • Jev calls require consent to send selected text and incur provider charges.
  • Keys belong in the environment, never chat, command logs or reports.
  • Website edits, baseline acceptance and report publication are separate actions.
  • Run the same audit again after approved changes.

Gate only new findings

anchorlint audit ./dist   --base-url https://example.com   --baseline anchorlint-baseline.json   --fail-on warning --format json   --output audit.json

A baseline must be reviewed. Regenerating it on every CI failure would hide regressions. Operational failures still exit 2.

Integrate without a wrapper

Claude Code, Codex and other shell-capable agents use this same workflow. The repository includes a portable SKILL.md; install it intentionally through your client’s documented skill workflow. AnchorLint does not modify agent configuration and is not an MCP server.

Agent guide ↗ · Portable skill ↗ · CI guide ↗