Skip to content

v0.1.0

Choose a tag to compare

@antondanv antondanv released this 30 Sep 09:06
· 63 commits to main since this release

First public release.

Added

  • One API over Claude Code, Codex and Antigravity: status(), ask(), askAll(), start() / run(), models(), ping().
  • A closed list of events with human-readable summaries, secrets masked.
  • Steering a working agent with hint() (Claude Code, Antigravity) and stopping it with stop(), an AbortSignal or a timeout.
  • Session resume, model and effort checked against each CLI's own catalog, MCP servers per run.
  • Access levels full, workspace and readonly, mapped to each CLI's permission and sandbox flags and verified with real runs.
  • Typed failures: usage_limit with the reset time, rate_limited, not_logged_in, network, timeout, stopped, empty_answer, failed.
  • The brainyard command: status, models, ask, run, ui.
  • A local dashboard that looks like a terminal, with status cards and a playground, over a token-guarded HTTP API with server-sent events.
  • Tests against fake CLIs, and npm run live for checks against the real ones.