Watch Codex work without watching its terminal
Codex sessions appear in the notch as their first activity arrives, labeled by project, with live state, current activity, files touched, diff counts, and a timeline of shell commands and patches. Subagents show under their parent, and per-session token counts are read back from Codex's local session log.
Approvals and questions, one keystroke each
When Codex asks a human to approve a command, the prompt surfaces in the notch with the command laid out: allow with ⌘ A, deny with ⌘ N, or deny with feedback explaining what to do instead. Codex questions arrive the same way, and the composer sends follow-up prompts into any resumable session. Auto-approved runs pass through without interrupting you.
The same window Codex meters you by
AgentPeek shows Codex's 7-day usage window as a live gauge with its reset countdown, read from what Codex reports locally, next to Claude Code's 5-hour and 7-day windows for an honest side-by-side. The by-day chart covers your last two weeks, and a budget alert fires before a window closes on you.
Setup without rewriting feature preferences
AgentPeek writes its Codex hooks to ~/.codex/hooks.json. Current Codex hooks are stable and default-on, so AgentPeek leaves feature preferences in ~/.codex/config.toml untouched. Settings, then Doctor, checks the managed wiring and hook trust. The Codex desktop app reads hook config at launch, so restart it once after install. Quick Routes jumps straight to the Codex config, skills, and session logs.