Stop watching AI work.

• Hooks are the source of truth

  State comes from the agents' own official lifecycle events.

• Processes are only for facts

  Discovery, enrichment, and real death. Never state.

• Unknown stays unknown

  A quiet session is never announced as finished.

• The allowlist lives in the hook

  It walks past everything not on it, without ever copying those bytes.

• Your prompts cannot leak

  Not to the socket, the logs, or the crash reporter. A test asserts it.

• 225 KB, no Foundation

  The hook has no general-purpose parser to leak through.

Claude Code / Codex
        │  official lifecycle hook (async, never blocking)
        ▼
codestatus-hook          ← metadata allowlist applied here
        │  one NDJSON line over a Unix domain socket
        ▼
CodeStatus daemon
        │
        ├── StateReducer      pure, idempotent, out-of-order tolerant
        ├── SessionRegistry   the single source of truth
        └── ProcessWatcher    kqueue NOTE_EXIT — discovery and death
                │
                ▼
        HUD · sound · notification

Codex input

No Notification event exists to observe it.

Codex in VS Code

Runs as app-server, and delivers the same hooks the CLI does.

Send prompt

Only for sessions we launch ourselves, through a PTY.

• How do I get notified when Claude Code finishes?

  Install CodeStatus and it registers a hook with Claude Code during onboarding. When a turn ends you get a macOS notification naming the repository, and the menu bar count moves from busy to free. Nothing to configure per project.

• Can I track several coding agents at once?

  That is the reason it exists. Every Claude Code and Codex session on the Mac appears in one list with its own state, so you can start four and keep working instead of tabbing between terminals to check on them.

• Does it work with Codex, including inside VS Code?

  Yes, both the Codex CLI and Codex in VS Code, which runs as an app-server and delivers the same hooks. Codex requires you to approve hooks once with the /hooks command before it will run them — a security control CodeStatus deliberately does not automate.

• How does it know an agent is busy rather than just quiet?

  It reads the agents' own lifecycle hooks, so state is reported rather than inferred. CPU is never an input: an agent waiting on a network response burns none, and one running a six-minute build burns plenty. When nothing has been reported, the session reads as unknown instead of being announced as finished.

• Does it send my prompts or code anywhere?

  No. There is no network code in the product at all — no account, no server, no telemetry. The hook copies an allowlist of metadata such as session id, event name and working directory, and walks past everything else without ever reading those bytes.

• Does it support Cursor, Windsurf, or Gemini CLI?

  Not today. It supports Claude Code and Codex, because it depends on official lifecycle hooks and those are the agents that expose them in the form it needs.

• How much does it cost?

  Nothing. It is free and MIT licensed, and the source is on GitHub.