Setup

Otty doesn't run agents for you — you keep using Claude Code, Codex, or OpenCode exactly as you do today. The one-time setup here teaches each agent to report its state back to Otty, so the window can show live badges, fire notifications, resume sessions, and surface history. It's a single click per agent.

Install Code Agent Integration

Open Settings (⌘,) and select the Agents tab in the sidebar.

Each agent has its own card. For every agent you actually use, click Install in its card. The button flips to Installed and a green ✓ Installed appears under Status.

What each install writes — and nothing else in those files is touched:

Agent	Button	Writes to
Claude Code	Install Hooks	~/.claude/settings.json
Codex	Install Hooks	~/.codex/hooks.json
OpenCode	Install Plugin	~/.config/opencode/plugins/

Restart the agent once

The integration connects only after the agent starts with the new config. If an agent was already running in a tab, restart it. For OpenCode, send one message so the plugin reports the session. For Codex, if you recently updated Otty or Codex, approve the trust prompt when it restarts.

You can click Uninstall in the same card at any time — it cleanly removes Otty's entries and leaves the rest of the file untouched.

What the integration does

Once at least one agent's hooks are installed, the Agent Behavior section at the top of the Agents tab comes alive. (The toggles are greyed out until then — there's nothing for them to act on.)

Setting	Default	What it does
Badge While Processing	On	Marks the tab while the agent is working on a task.
Badge When Task Completes	On	Marks the tab when the agent finishes.
Badge When Awaiting Input	On	Marks the tab when the agent is waiting for approval or input.
Notify When Task Completes	On	Sends a system notification on completion.
Notify When Awaiting Input	On	Sends a system notification when the agent needs you.
Prevent Sleep While Processing	Off	Keeps macOS awake while a task is running.
Resume Session on Recovery	On	Reopens the agent session when a recovered terminal comes back.

These state signals are what power the rest of the agent experience: the tab badges and dashboard in Monitor Tasks, the live transcript and Resume in History, and the branch / fork actions for spinning a new task off an existing context.

Usage readout

Open the details panel (⌘I) on the Info tab while an agent is running and its group shows what the session has spent: how much of the context window it is using and its running token total. There is nothing to turn on — the numbers are read straight out of the log that agent already writes, so it costs a file read and nothing leaves your machine.

What each agent reports differs, because each one records something different:

Agent	Plan quota	Context	Tokens	Cost
Claude Code	On by default	Yes	Yes	—
Codex	On by default	Yes	Yes	—
omp	On by default	Yes	Yes	Yes
OpenCode	—	Yes	Yes	Yes
pi	—	Yes	Yes	Yes
Kimi Code	—	Yes	Yes	—
Cursor CLI, Grok Build, Reasonix	—	—	—	—

Bars need a real ceiling. Codex is the one agent that states one: its context window goes into its own session log, and its plan window comes back from the server, so both of its readings get a bar — amber past 75%, red past 90%, warning before a limit bites rather than reporting it after. Every other agent would need that ceiling guessed from a model id, so they show the measured occupancy on its own. Claude in particular records nothing that distinguishes a 200k session from a 1M one.

Codex's quota is asked for, not overheard. It used to be read out of the rate_limits block in its session log, which was free but frozen at whatever that session's last request saw — a few idle days here and the panel showed days-old figures, and spend from another machine never appeared at all. It now asks codex app-server, which answers from the server.

Plan quota is reported by three agents — Claude Code, Codex and omp — and all three show it out of the box. To switch one off, expand its card under Settings (⌘,) → Agents and turn off Show Plan Quota; the setting records the agents you turned off, so an agent that learns to report a quota in a later version arrives switched on rather than silently dark.

• Claude Code shows Session quota and Weekly quota bars, each with its reset time underneath — the same two numbers /usage shows you inside Claude Code. It comes from claude -p "/usage".
• Codex shows its plan's window plus any per-model allowance as separate bars, since they are separate budgets. It asks codex app-server directly, so the figures are live and include usage from your other machines.
• omp gains its 5-hour, weekly and monthly spend limits, from omp usage --json. Every account you have signed in is pooled into one bar, the way omp's own capacity line counts them — rotating between accounts is the point, so a single account's ceiling would understate what you have. The figures cover spend omp itself brokered; anything you spent through another client is invisible to it, which is what the As observed by omp line means.

These have a switch at all because they are the only readings that cost anything: each runs that agent's CLI in the background — the same command you'd type yourself — which takes a couple of seconds rather than a file read. If the CLI isn't installed, Otty stops trying after a few failed attempts rather than retrying forever.

The other agents report no limits at all, so there is no switch on their cards.

It re-reads every 5 minutes, and again whenever a task finishes, since that is when the numbers actually moved. Several agents finishing at once still costs one reading. Claude's probe runs on the cheapest model, at low effort, and with session persistence off, so it neither shows up in your session history nor costs anything meaningful. Every one of them uses the login that CLI already has: no sign-in, no stored password, and nothing read from your Keychain.

The toggle is also in the command palette and the workspace agent menu as Show Plan Quota.

Cost is only ever the agent's own number. OpenCode, pi, and omp compute what a session spent and write it down, so that figure is shown as-is. The others log tokens but no cost, and pricing those tokens against a public rate card would produce an API-equivalent number that no subscription is actually billed at — so nothing is shown rather than something misleading.

Skills

At the bottom of the Agents tab, the Skills card hands your coding agent a few ready-made abilities that drive Otty from inside a session:

Skill	What the agent can then do
/wait	Block until something finishes — a command, another agent's session, or a pane / tab / window going idle.
/tell	Send a message to another terminal: an agent session, a window, a tab, or the active pane.
/open	Open a file or URL inside Otty — in a split, a new tab, or a new window.

Click Copy on a row, then paste into your agent — it writes the skill file itself. See Skills for what to type afterwards, and for the permission /tell needs before it can type into a pane.

See also

• Composer — the input affordance for agent sessions.
• History — search and resume past sessions.
• Monitor Tasks — running multiple agents at once.