Cue

One window for every coding agent that needs you. When a Claude Code, Codex or Pi session finishes its turn, asks permission, or asks a question, it shows up in Cue — so you stop cycling through tabs to find who's waiting. A Mac app (macOS 13 or later, Apple silicon and Intel).

• Finished, your turn — see the agent's last message and reply from Cue: it types your message into that session's terminal (iTerm or tmux) and presses Enter. Or jump to the tab.
• Permission — Allow, Deny, Always allow, or "tell it what to do instead".
• Question — pick an option (1–4) or type your own answer.

The terminal keeps working too: answer there or in Cue, whichever is first wins, and Cue clears the card either way.

Also in Cue

• Actions — your project's own buttons (build, test, start a dev server) under the agent's latest reply: click one instead of asking the agent to run it. Two or three per project, kept in Cue on your Mac (never in the repository); a new or changed command runs only once you've allowed it. See docs/actions.md.
• Talk to it — hold right ⇧ in any app, say what you want, and let go: it goes to the chat open in Cue (else the session that moved last). When the answer is back, Cue's island at the top of the screen says so, small, and waits for you: Listen reads it out, Read shows it there, Later puts it away. Live captions can show both sides along the bottom of the screen instead, the answer waiting as a small tab in its corner. Off until you turn it on (Settings → Voice). What you say is written down on your Mac, by Apple's recognizer or, if you pick it, NVIDIA's Parakeet (Apple silicon), so nothing you say leaves it.
• Listen and Read — Listen reads a reply out loud, in your Mac's voice or a natural voice that runs on your Mac (Settings → Voice); Read opens a long reply full width.
• Insights — what your sessions cost, by day and by model, and the part that was avoidable: caches written again after a session waited on you, and long conversations read again past their hand-off point. A waiting session counts down its cache and can write its hand-off note before it runs out, or start fresh from its last reply once it has.
• Steps, Diff and background agents — what a session read, edited and ran between your message and its reply; what it changed; and the agents it sent off, each at work under it.
• Projects — sessions group by repository, with a color and a pin each; move a session to another project from its ⋯ menu (Settings → General → Reset projects… undoes every change).

Install

1. Download Cue-<version>.dmg from the latest release, open it, and drag Cue to Applications. It's signed and notarized by Apple.
2. Open Cue and click Connect Claude Code (Settings has Connect for Codex and Pi too). That adds Cue's hooks to ~/.claude/settings.json; the file is backed up first and your other settings stay. In a session that was already open, type /hooks once (or restart it; claude --resume keeps the conversation).
3. The first time Cue jumps to or types into a tab, macOS asks to let Cue control iTerm or Terminal: allow it.

Cue updates itself: it checks for a new version now and then, and the menu bar icon's menu has Check for Updates….

From source instead: ./install.sh builds the app and connects Claude Code, Codex and Pi; ./install.sh --uninstall removes it all again.

Your own usage meters

Behind a proxy or on a team budget? Keep a small CSV up to date (a cron job, a script, whatever you like) and Cue shows it in the usage pill, beside Claude Code's own limits. The default path is ~/.cue/usage.csv (Settings → Usage file changes it):

label,spent,limit,unit,resets_at
Team,42.60,50,$,2026-10-12T00:00:00Z
Credits,1200,,credits,

Only spent is required. An empty limit shows the amount without a bar; unit defaults to $; resets_at is an ISO date-time or Unix milliseconds. Up to 3 rows. Cue only reads the file, whenever it changes, and never runs anything. If it doesn't match this format, the pill says what's wrong.

Extensions

An extension brings a page of its own, as a tab beside a session's chat or in the Extensions panel (the button beside the terminal's). Cue comes with one, Diff: what the session changed this turn, the whole session, or everything not committed. Switch it on in the Extensions panel (the button beside the terminal's in the top bar); it updates with Cue. To add another, copy its folder in from there (Open the folder). Nothing in an extension runs until you switch it on. To make one, see docs/extensions.md.

Privacy

Cue collects nothing and sends nothing anywhere. Everything it keeps (what's waiting, history, drafts, settings) stays on your Mac, in the folder below. It talks to your agents through a local socket that only processes on your Mac can reach, and it reads their transcripts on your disk. The one time it goes online is to check GitHub for a newer version of Cue, and, only when you pick them, to download a voice model once (a natural voice for Listen, or NVIDIA's listening model for talking to it); after that they run on your Mac, and what you say never leaves it.

An extension you switch on can't reach the network. It can have Cue read git, read-only, in the open session's folder. It can run programs only if it asks to and you approve it: then only the installed programs it lists, never one it brings, a shell, or a script interpreter. Settings shows what each one may do and every command it ran.

Using it

Key
A / D	allow / deny
1–4	pick an answer
R	reply / redirect (type, then Enter)
G	go to the session's tab
← →	flip through what's waiting
H	history

Everything Cue keeps is in ~/Library/Application Support/dev.spacegrowth.cue/ (the standard place for a Mac app's data; CUE_HOME overrides it), in one SQLite file, cue.db, so a restart or crash loses nothing:

Table	What
items	decisions and finished turns, pending and answered (the history)
sessions	every session, live or ended, with who drives it (driven_by)
exchanges	each session's full back-and-forth, append-only
drafts	what you typed and haven't sent, per session

Images you send stay as files in uploads/ there (rows hold their paths).

Settings stay in config.json there (the Pi extension reads it). How much history to keep is set in Settings or there:

{ "history": { "keep": 300 }, "pi": { "gate": "dangerous" } }

Pi has no permission prompts of its own; Cue's extension adds them for risky commands — pi.gate is "dangerous" (default), "all" (every bash/write/edit), or "off".

Agents that run other agents

If a tool starts agent sessions on behalf of another agent (a lead handing work to helpers), it can set CUE_DRIVEN_BY in each helper's environment, e.g. CUE_DRIVEN_BY="its lead (billing)". Cue then keeps that helper out of Waiting: its finished turns are the lead's to handle, so it shows as idle, "waiting on its lead", with no notification. (Settings can show them anyway.)

How it works

Claude Code ──hook──┐                       ┌── window (spotlight card)
                    ├──► cue.sock ─────────► Cue ──┤
Pi ──extension──────┘    (one JSON line)    └── notification

Each harness has a hook that waits for an answer: Claude's PermissionRequest, Pi's tool_call. The hook asks Cue over a local Unix socket and waits. Turn-end events (Stop, agent_end) put a "finished" card up; the next prompt (UserPromptSubmit, agent_start) takes it down.

• Cue not running? Hooks exit silently and the agent asks in its terminal as usual.
• Answered in the terminal? Claude doesn't tell the hook, so Cue watches the session transcript and clears the card. Pi's extension tells Cue directly.
• Cue restarted? Waiting hooks reconnect and ask again.

The hook is built into Cue itself (cue hook <event> <harness>, see src-tauri/src/hook.rs), so it needs nothing installed: no Python, no Node. Adding another harness means one small hook that speaks the same socket protocol.

Develop

npm install
npm run dev                      # run the app
./tests/run.sh                   # all tests, against a private silent Cue (yours keeps running)
python3 -m http.server 8765      # then open /tests/preview.html — the real UI with sample data
hooks/cue-ctl list               # poke a running Cue from the shell (needs python3)

License

MIT. See LICENSE.