Every session, live
Run claude in any terminal and the session appears in the notch as supported activity arrives: project name, state (executing, thinking, waiting, or idle), current activity, files touched, commands run, and diff counts. Expand it for the transcript, the timeline of tool calls and prompts, Claude's own todo checklist updating as it works, subagents under their parent, and model, account, and git branch chips, with a context-limit warning before compaction takes you by surprise.
Answer Claude without finding its terminal
When Claude Code asks permission for a command, edit, or tool call, the prompt lands in the notch with the command or diff laid out: ⌘ A allows, ⌘ N denies, ⌥ A covers Always Allow when offered, and deny with feedback sends a note back. Questions and plans get the same treatment, and the composer sends the next prompt into any resumable session with the reply streaming in. The full flow is on the permissions page.
The 5-hour and 7-day windows, always visible
AgentPeek reads the usage Claude Code itself reports on your Mac and shows both rate-limit windows as live gauges with reset countdowns, plus a by-day chart and per-session token counts read from the session transcript. A budget alert fires when reported usage crosses your chosen threshold (50 to 100 percent), so a long-running refactor wraps up instead of stopping mid-file.
Managed setup
AgentPeek installs its Claude Code hooks into ~/.claude/settings.json on first launch, and Settings, then Doctor, repairs them if a CLI update ever disconnects them. Quick Routes opens skills, agents, commands, rules, plugins, the settings file, and session logs, one click each.