A local command center for every AI agent you have in flight. Claude Code sessions and Codex threads, in one glance, so you always know which one needs you right now.
As agents do more work unattended, the hard part stops being doing the work and becomes knowing which of your N running things is waiting on a decision. Baton reads the signals your machine already emits (zero manual entry) and surfaces the one that matters: the 🎽 "the baton's with you" bucket, an agent that ran its leg and handed back to you. That's the thing that's easy to drop when you're looking elsewhere.
It lives in your macOS menu bar (a relay runner + "N batons for you"); click for the full picture, click any session to jump straight to it.
curl -fsSL https://raw.githubusercontent.com/neilkpatel/baton/main/install.sh | bash
That clones Baton into ~/.baton
, sets up an isolated Python venv (the only deps are rumps
and the FSEvents binding), and installs a login LaunchAgent. The 🎽 appears in your menu bar immediately, starts at login, and relaunches if it crashes.
Prefer to look first? Clone and run the same script from the checkout:
git clone https://github.com/neilkpatel/baton.git && cd baton
bash install.sh # or: bash install.sh --no-autostart
Update: re-run the same one-liner. It pulls the latest into ~/.baton
and restarts the app.
Remove the autostart anytime with bash install.sh --uninstall
.
Requirements: macOS, Python 3.9+ (3.11+ enables Codex scheduled-automation tracking), and Claude Code and/or Codex. Baton shows whatever it finds; either alone is fine.
Menu bar glance: the title is a live count of sessions waiting on you. Always visible, updates inunder a second(itwatchesyour session files with FSEvents rather than polling, so it's both instant and battery-cheap; it idles until something actually changes).Two agents, one view: the dropdown groups what's waiting by tool (Claude Code / Codex), then Working, then Done.Click to jump: click a Claude session and itsTerminal.app tab comes to the front; click a Codex thread and it opens inCodex via itscodex://
deep link. No more hunting through windows.Accurate "waiting" signals, per tool:**Claude Code: a session that's idle with an assistant answer as its last turn = waiting.Codex: mirrors Codex's ownunread (blue-dot)** state, so a thread you've already opened doesn't nag you.
Click-to-acknowledge: jumping to a waiting session clears it from the count until it produces anewanswer, and the session you currently have open in Terminal isn't counted while you're looking at it. The number goes down when you deal with things.Optional hand-off notifications: off by default (the menu bar is the calm channel); one toggle turns on a banner the moment a baton comes back.Recognizable labels: each session is titled by Claude Code's own running summary (ai-title
), so you know what it's about at a glance.
There's also a full HTML dashboard (server.py
index.html
) for the rich multi-tab view. Open it from the dropdown ("Open full dashboard →"), or:
bash start.sh # python3 server.py --port 8787 → http://127.0.0.1:8787
Every source is a collector that normalizes into one track
record, so the UI never changes as sources are added:
{ id, source, title, project, status, lastActive, detail, alive, extras }
status ∈ waiting | working | done | idle | scheduled
| Source | Read from | Signal |
|---|---|---|
| Claude Code | ||
~/.claude/sessions/{pid}.json + transcripts |
||
live busy /idle + last turn → working vs waiting-on-you |
||
| Codex threads | ||
~/.codex/session_index.jsonl + .codex-global-state.json |
||
| unread (blue-dot) → waiting; rollout transcript → the agent's closing line | ||
| Codex automations | ||
~/.codex/automations/*/automation.toml |
||
| scheduled pipelines → next/last run |
Claude Code (~/.claude
) and Codex (~/.codex
) are separate tools with separate state dirs; Baton reads both independently. Detection is event-driven (FSEvents on the session dir) with a slow safety-net poll.
Single-file collectors + menu bar app in stdlib Python (plus rumps
for the menu bar and
pyobjc-framework-FSEvents
for the watcher, in an isolated .venv
). The dashboard is dependency-free vanilla HTML/CSS/JS. No build step, no framework.
Click-to-jump raises Terminal.app tabs only. If you run Claude Code in iTerm, Ghostty, or another terminal, detection and counts still work fine; clicking a Claude session just can't raise your tab yet (it opens the dashboard instead). Codex jumps work regardless, via deep link.Claude Code has no native read/unread flag(Codex does, and Baton mirrors it directly). For Claude, "waiting" clears when you view the session's Terminal tab, click it in Baton, or hit "Mark all as seen", and it comes back when the session produces a new answer.macOS only by nature: FSEvents watcher, menu bar app, AppleScript jump.
Entirely local and strictly read-only over ~/.claude
and ~/.codex
: Baton never mutates your session state, and nothing leaves your machine. The live payload contains your real prompts and working directories, so the dashboard server binds to 127.0.0.1 only, never
0.0.0.0
.
Captured state (state.json
) is git-ignored; only scrubbed sample data is committed.menubar.py
(the app) · collectors.py
(the signal readers) · server.py
index.html
(dashboard) · install.sh
· start.sh
· requirements.txt
· state.example.json
(the
track
contract).