Skip to main content
Install and first steps are in the Quickstart. self-bench help lists every command in your terminal.

Sign In and CI

login saves a personal API key to ~/.config/self-bench/config.json, readable only by you. On a remote machine, paste the address the browser lands on into the login page. logout revokes the key. In CI, skip login and set SELFBENCH_API_KEY (from Settings → API Keys) and SELFBENCH_ORG. SELFBENCH_API_URL or login --api-url points at a self-hosted deployment.

Conventions

  • Workspace. Commands act in the workspace login chose. org use changes it; --org overrides it once.
  • Scope. R in the command tables is --repo OWNER/NAME. An eval is evals/NAME. --benchmark ORG/NAME and --group SLUG take its place.
  • Tasks are named by batch and ID: RUN_ID TASK_ID.
  • Configurations are named by the KEY that results prints.
  • Output. Tables on a terminal, JSON when piped or with --json, so self-bench results --repo acme/app | jq works.
  • Background work. Explanations, audits, and summaries return at once. Add --wait to wait and print the answer.

Commands