replay-theater

<img src="docs/assets/app-icon.png?v=2026-09-23" width="42" alt="" align="top"> AgentBar

One approval queue for your coding agents, and rules that watch before they answer.

<img src="docs/assets/agentbar-tour.gif?v=2026-10-16" width="640" alt="AgentBar tour: the island under the notch opens on a Claude permission request for git push origin main with Codex and Copilot sessions listed below, one click on Allow; then the welcome window's Menu bar / Dynamic Island / Both choice, then a walk through Settings">

Why. Your agents stop and ask before they run something, and the question waits in whichever terminal tab they happen to be in. AgentBar is where they ask — every agent, every session, in the menu bar or under the notch — so you answer in a click instead of hunting for the window. Nothing is answered behind your back: only you do, or a rule you wrote yourself, and every answer is written down.

Things to try

What it covers. Agents: Claude Code and Claude Cowork, Codex, Cursor CLI, Gemini CLI, GitHub Copilot, Google Antigravity, Qwen Code and OpenCode — plus Cursor cloud agents, Devin and Codex cloud tasks, and sessions on your own machines over ssh. Terminals: any — the exact tab in iTerm2, Terminal, WezTerm and tmux, best effort in kitty, Ghostty and VS Code, Cursor or Zed, and anything else (Warp included) comes forward as the app the agent runs in. Sessions: as many at once as you run, the one that needs you first. Approvals: a per-repo history of what you decided, and rules you wrote, every firing on record. Surfaces: the menu bar, a Dynamic Island under the notch, or both, pinned to the display you pick. Elsewhere: the agentbar CLI on Linux, and AgentBar for Windows, a native system-tray counterpart on the same ~/.agentbar hook protocol.

Quick start

curl -fsSL https://raw.githubusercontent.com/michalstrnadel/AgentBar/main/Scripts/install.sh | bash

1. The app lands in /Applications (or ~/Applications when that isn't writable; AGENTBAR_INSTALL_DIR overrides), launches, and installs its hooks. A welcome window asks where it should live — menu bar, Dynamic Island, or both.
2. Open a new Claude Code session (hooks load at session start) and give it any task.
3. Watch it: the mascot animates while the agent works, and the moment it asks for permission you get a needs approval row — click ✓ Allow, ✓ Always, or ✕ Deny right there. No terminal switch needed.

That's the whole loop. More install options below; troubleshooting at the bottom.

What the installer changes (and how to undo it)

AgentBar is local-only — no telemetry, and it talks to three places, two of which stay silent until you switch them on: the update check every few hours against GitHub Releases, and — only if you tick Settings ▸ Usage — Claude's own quota, from api.anthropic.com using the login Claude Code already stored, or from claude.ai if you sign in there. That switch reports what came back, in a sentence under itself, including every way it can fail; Check now asks again on the spot. The simplest way in is Sign in to Claude…, which opens claude.ai's own login page in a window and keeps the session in AgentBar — no terminal, no token, and no reading of your browser's cookies, which AgentBar does not do. If your sessions run under their own CLAUDE_CONFIG_DIR, that login is somewhere AgentBar cannot read — Use a token… takes one from claude setup-token and keeps it in AgentBar's own Keychain item, which the same button removes. Everything else, including every token count and every other provider's quota, is read from files already on your disk. The install touches exactly these, all reversible (see Uninstall):

• Copies the hook scripts to ~/.agentbar/hooks/.
• Merges AgentBar hook entries into your Claude Code settings — ~/.claude/settings.json, and your CLAUDE_CONFIG_DIR if you set one. Existing hooks are preserved.
• Writes into ~/.codex/config.toml only if you use Codex: a notify line if it has none, and a hooks block between two marker comments — what is outside them stays as you wrote it, and Codex asks you once before running any of it. Merges into ~/.cursor/hooks.json / ~/.gemini/settings.json / ~/.gemini/antigravity{,-cli}/hooks.json / ~/.qwen/settings.json only if those exist. Writes its own ~/.copilot/hooks/agentbar.json only if you use Copilot — a separate file, so your own hooks stay untouched.
• Copies a plugin to ~/.config/opencode/plugins/agentbar.js only if you use OpenCode.
• Copies the Claude Code mod to ~/.agentbar/mods/ — and that is all, until you switch it on: it is the one integration that starts off.
• The SessionStart hook launches AgentBar in the background when an agent session begins.
• Before it writes into any of those settings files it keeps the file as it was, beside it — settings.json.agentbar-bak-20261001-142233 (local time) — and keeps only the newest three of its own copies per file. A launch that would change nothing writes nothing and keeps nothing, so the copies only appear when something actually moved. Settings ▸ Agents ▸ Show changes… (or See what changed… in the welcome window) shows every write as a unified diff with the copy it kept, and what a re-install would change right now before you let it. On Linux, agentbar install-hooks prints the same diff for each file before writing it and keeps the same copies.
• Every agent it finds is wired, unless you turned it off: the switch in Settings ▸ Agents, agentbar unwire <id> on Linux (or install-hooks --skip a,b / --only a,b). Turning one off takes AgentBar's entries back out of that agent's settings — backed up and diffed like any other write — and every later run leaves it alone (~/.agentbar/wire-disabled). agentbar wire <id> or the switch puts it back.
• Nothing else is granted automatically: the exact-tab jump-back asks for Automation access the first time you click a row, and approving a plan (or a Codex/Copilot prompt) asks for Accessibility. Decline either and AgentBar falls back to bringing the app forward and letting you answer there.

The installer prints this summary before doing anything, and never modifies a tool you don't use. To undo a single change, copy the .agentbar-bak-… file back over the original. Hooks are snapshotted per session — start a new agent session afterward.

Features

<img src="docs/assets/demo-claude-codex.gif?v=2026-09-23" width="640" alt="AgentBar demo: Claude session works, needs approval, one-click Allow, then a Codex session takes over the bar">

<img src="docs/assets/demo-island.gif?v=2026-09-23" width="640" alt="AgentBar as a Dynamic Island: the pill under the notch says approve?, opens on hover into the session panel with the mini-diff, one click on Allow, and the pill flashes ✓ Allowed">

<img src="docs/assets/hand-a-file.gif?v=2026-10-07" width="720" alt="A screenshot dragged from the desktop onto a Claude session in the island: the row says 'Drop to hand it to Claude', then '✓ In Claude's prompt — add a word and press Return', and the escaped path appears in the terminal's prompt with no Return pressed. A Codex row says 'quiet 12m?', and the footer meter reads 'Claude 12% left · out ~15:03' in amber">

• Two surfaces, your pick — the classic menu bar item, a Dynamic Island pill under the notch, or both. Pick it in the welcome window on first launch, change it any time from Appearance…; no relaunch. See Dynamic Island.
• Live status per agent — an animated mascot works the bar while an agent works: Clawd the crab (Claude), who acts out what the session is doing in twelve scenes, the knot + a braille dot-matrix that literally spells codex (Codex), the pixel mascot head + dots spelling copilot (Copilot), and the animated pixel rainbow arch (Antigravity).
• Notifications that only carry what wants you — off by default, and three separate switches: an agent needs approval (Allow and Deny on the banner itself, answerable without switching apps), a session failed, and everything went quiet — one summary of the whole batch, and only once you have actually been away from the keyboard for a couple of minutes. Nothing is announced for merely finishing. Send a test checks they reach you; if nothing appears but it shows up in Notification Center, a Focus is on — macOS files banners rather than showing them, which is working as designed.
• What happened today, with its weight — sessions that finished, how long they actually worked (not how long their windows were open), what they cost and what moved in the repo: "12 sessions · 3h 40m · 4.1M tokens", and per row "AgentBar · 34m · 1.2M · 7 files +210 −80". Token counts come from each agent's own local files (Claude's transcript, Codex's rollout, Copilot's session store); the agents that publish nothing simply show nothing, never a zero. In the menu bar dropdown it is the Today row, and behind the island's ⋯. Appearance… can also put it along the bottom of the island — the day's total, a bar per session, or both. agentbar history says the same in a terminal. Live status forgets a session the moment it ends; this doesn't.
• Pick the island's display — on a multi-monitor desk the island can be pinned to one screen instead of following the pointer around. The welcome window draws your displays the way System Settings does; unplug the pinned one and it falls back to the pointer until it's back.
• Permission alerts — an amber dot the moment an agent waits for your approval.
• It remembers what you decided — the fifth time an agent asks to run the same thing, the card says so: “Allowed 23× here · last Tue”, counted per repo, because a command that is routine in one checkout is the opposite in another. Once a prompt has been allowed five times and never refused, ✓ Always is pointed at — pointed at, never pressed. Kept in ~/.agentbar/decisions.jsonl, never sent anywhere, switchable off in Settings ▸ Approvals, and agentbar forget empties it. Keystroke approvals (Antigravity, and Codex sessions older than its hooks) write nothing: a key pressed at a terminal is not a decision anybody here witnessed.
• Rules you wrote — the one thing AgentBar will answer without asking, and only ever a rule you typed yourself. When you have answered the same prompt the same way five times, the card offers to write it down; Settings ▸ Rules is where they live, in plain JSON at ~/.agentbar/rules.json that you can edit by hand. Writing one shows you what it will not answer, and gives you a field where you type a real command — git push --force origin main — and are told on the spot whether this rule would have taken it, and which clause stopped it. A new rule starts out watching: it answers nothing and writes down what it would have done, so you can look at a week of that before you let it speak for you. The rule's line says whether you, answering those same prompts, did what it would have done — and once it has matched you ten times across three days without a single disagreement, it offers a Let it answer button that still asks before it changes anything. A rule that refuses may cover every repository on the machine. A rule that approves names one — and before it answers, the command itself is checked again, not just its shape: anything chained, piped, redirected or substituted, anything under sudo, a destructive git or rm, anything reaching off this Mac, a path outside that directory, or anything that touches how permission itself is configured comes back to you. No setting turns that off. Every firing writes a row naming the rule, so Settings ▸ Rules and agentbar rules can tell you what each one has actually done. Nothing in the file applies while any of it is wrong, and Diagnostics says so — because a rule that silently stopped working looks exactly like AgentBar working normally.
• Your week of decisions — the top of Settings ▸ Approvals lists the five prompts that held your agents up longest over the last seven days: how often you were asked, how long they waited (“14× · 7m waited”, or “at least” when an older hook left a wait unrecorded), and how you answered. Beside each one is the rule you already wrote for it — and while it is watching, what it would have done: “would have answered 9 of 9, about 4m”. A prompt with no rule that you answered the same way five times offers Write a rule…, which opens the ordinary rule sheet filled in; the rule starts out watching and is saved only when you press Add rule. Only your own answers in AgentBar are counted.

<img src="docs/assets/week-of-decisions.png" width="560" alt="Settings ▸ Approvals, Your week of decisions: since 1 Oct, asked 49×, 25m waited, your rules answered 5 more. git push 14× · 7m waited, allowed 14, no rule yet, with a Write a rule… button; rm 6× denied, its deny rule is off; npm test 11×, a watching allow rule would have answered 9 of 9, about 4m; edit Sources/*.swift 11×; curl 3×">

• How long they waited on you — the other half of the day's account, under Today: “18 answered · 3 by your rules · they waited 34m on you”. Nothing else on the machine is standing in the right place to measure it, and the two counts stay apart — a rule's answer is not one you gave. agentbar approvals says the same, with the prompts you answer most.
• Multi-session — every running session listed with its agent's mark, project, git branch, state and elapsed time; click a row to jump to its app or terminal.
• Open anything — launch any supported agent (Claude, Codex, Copilot, Antigravity, Cursor, Gemini, Qwen, OpenCode) straight from the menu.
• Start a task, not just an agent — New Task… in the menu (or ⌥⌘N, once you switch that on) opens a small panel: a project you have worked in, an agent, and a line of what you want. The agent opens in a terminal, in that directory, with the prompt already given. It closes the moment it loses focus and takes no space until you ask for it. The prompt goes in as an argument, never as synthesized keystrokes, and only to the agents whose CLI documents one — the rest open in the right place and wait for you to type. A terminal that can't be handed a command (Warp) gets the command on your clipboard and says so, rather than opening on the wrong thing.
• Two looks — full-color mascots, or a monochrome System mode that matches the menu bar.
• Remote Allow/Deny — answer Claude Code permission prompts straight from the menu: see exactly what's requested, then Allow once, Always allow, Deny, or defer to terminal.
• A diff you can actually read — an edit shows the lines that moved, with a line of context and a marker where untouched lines were skipped. Where a single line was edited, the characters that changed stay at full strength and the rest of the line fades, and a change past the right edge slides into view instead of truncating — both halves by the same amount, so the columns still line up. The +N −M beside it counts what moved, not the size of the window it moved in.
• Answer questions too — when Claude asks a multiple-choice question, the island and the menu show the actual options: tap one and the session continues, no terminal switch. The terminal wizard stays live the whole time — whoever answers first wins. Multi-question calls become a one-question-at-a-time wizard on the island: each tap records and slides to the next, with Back and a 2/4 mark.
• Allow all alike — three agents waiting on the same npm test in the same folder are one decision: the card says Allow all 3. Only the exact same request counts (same tool, the whole same input, the same directory), only the ones on screen when you click, and each goes into the approval history as your own answer.
• Deny with a note — refuse and say what to do instead: "use pnpm here, not npm". Type it on the island card (Deny with a note…), on the notification banner, or agentbar deny --note "…"; the agent reads it as the reason and changes course rather than trying the next thing. On a plan it is the feedback the plan goes back with. A denying rule can carry one too (Tell it), so a refusal you wrote down once explains itself every time it fires.
• Plan review — when Claude finishes planning, the full plan renders on the island as formatted Markdown (scrollable when long). Keep planning sends Claude back to refine it without touching the terminal; Approve plan jumps to the session's exact tab and answers the plan dialog for you.
• What's left, at a glance — one small meter per provider, answering one question: how much is spent and how much is left. On the island it is the footer line, drawn rather than written, at the height that line always had; in the menu it is the fuller block, with both windows and their reset times. Codex reports the exact percentage of its 5-hour and weekly windows with real reset times, plus a credit balance when the account has one. Copilot carries its own priced ledger, so its line is what it actually charged today in its own AIU — and it gets no bar, because the ceiling lives on github.com and a meter drawn against a guessed one would be a picture of a number that doesn't exist. Claude keeps its windows on its own servers: tick Settings ▸ Usage and AgentBar asks for them with the login Claude Code already stored (off by default, five-minute polling, and it never touches your refresh token); leave it off and you get the tokens its transcripts record for the current 5-hour block. A window past its reset says so rather than repeating the old number, everything is hidden the moment it goes stale, and agentbar usage says the same in a terminal.
• Will it last? — AgentBar fits a line through the last 45 minutes of each quota window and, only when it says the window runs out before it resets, tells you when: "out ~15:40" on the island (amber inside the last half hour), "At this pace: limit ~15:40 (+18 %/h)" in the menu. A forecast that changes nothing says nothing, and none of it ever notifies.
• Carry it on elsewhere — when a session's quota is half an hour from running out, its row says out ~15:40 ↗; click it (or right-click any session row, or use Continue … Elsewhere in the menu) and pick another agent. The launcher opens in the same project with that agent and a prompt that says where the work got to — your last prompt, the agent's last update, and "look at git diff first". You read it, you press Return; the first session is left exactly as it was.
• Hand a file to an agent — drag files, an image from the browser, or a screenshot thumbnail up to the notch and drop it on a session: its path goes into that agent's prompt, escaped the way Terminal escapes a dragged file. It is pasted only into a tab AgentBar has verifie