Skip to content

Dashboard

arl ui                          # opens http://127.0.0.1:8765 over .arl/arl.db and arl.db
arl ui --db runs/ci.db --db arl.db
arl ui --export site/           # self-contained, read-only copy to attach or host on GitHub Pages

Every run is presented as an investigation file:

  • Dockets: every agent you've tested, its pass rate over time with 95% intervals, and each run with its source (manual cli, background check, mcp, ci).
  • Run docket: the verdict first ("8 problems found"), the pass rate with its interval, then the probable causes: each problem in plain words, why it matters, the exact attack and the agent's reply, the fix, and every scenario it affected. Below: pass rate by category, measurements (injection resistance, safety violations, recovery after tool faults, hallucination rate, loops, crashes, latency, tokens), what the run covered across the 8 capabilities and OWASP risks, and every scenario with filters.
  • Recorder transcript: one conversation as a sequence of events: user turns, tool calls with injected faults stamped, agent replies with their real durations, crashes, and every check with its verdict.
  • Compare: two runs side by side with the statistical test the CI gate uses; regressions are stamped.

States are printed marks, never faded rows: FAILED, FLAKY, ALWAYS FAIL, SKIPPED, N/A. Skipped or ungraded scenarios never look like passes.

Privacy

The server binds to 127.0.0.1 and is read-only. Transcripts contain attack text and your agent's replies, including anything it leaked; treat exports like logs. All content is escaped before display, so attack payloads such as <script> show as text.

No network requests: no CDN or analytics, so it works offline. The display face, D-DIN by Datto (SIL Open Font License 1.1), ships inside the package.