Operator

A Mac app for Claude Code. You run the agents: decide what each one does, then tune every lane’s model and effort so the plan runs close to its limit and never past it.

When2025 to now
ForIndependent
Built withReact, Vite, Tailwind, Electron, node-pty, xterm.js
LinksThe landing, source on GitHub (MIT)

The brief

I run several Claude Code sessions at once, often on the same repo. I decide what each agent does, hand work between them, read the diff before it lands, and tune each lane’s model and effort: cheap models where the work is mechanical, high effort only where it pays.

It is a living tool, not a product: no roadmap, no support queue, no telemetry, no account. If it is close but not quite right for you, the source is yours to change.

The mark

A field of dots pulling into a circle: many agents, one center.

The mark. The dock icon.

Orchestrating

Staff the project once, then work the board. Each lane pins a model, an effort level and permissions. The card is the task, the agent is a chip on it, and Waiting is the column that needs you.
Hand work between agents. Every hop lands in the dispatch log with its outcome. A dispatch from a lane that is not the coordinator waits for your approval, and nothing retries.
Watch the call tree as it happens, rebuilt from Claude Code’s own transcript files. Nothing to instrument.
Run agents in parallel: one git worktree per task, so nothing touches your working tree until you say so.

The team rail beside the board: lanes with their models, task cards in columns. — The team and the board.

The comms log: four dispatches between lanes, one held for approval with Approve and Decline buttons. — Comms. One dispatch held for approval.

The agents list: five agents, each with a model chip and an effort level. — Agents, each with its model and effort.

Tuning

Set the model and the effort per agent, in the .claude/agents files you already have. These two decide what a lane spends.
See what each model spends, per project and per day, read from the transcripts on disk. Tokens, never dollars.
Run close to the limit, never past it. Session, week and per-model caps, each with its reset. A timed-out read says no data, never 0%.
Review the diff and land it in place: commit, merge or discard.

The plan meter: session, week and Opus bars with their reset times, and API billing reading no data. — The plan meter. Absent is not zero.

The review panel: a diff of auth/session.ts with three lines added and one removed. — Review, before it lands.

Scope

Operator hosts Claude Code’s own CLI over a pty and reads the transcripts it writes. It stops an agent with Claude Code’s own interrupt, so the session survives, and passes effort as a launch flag, because settings.json drops the fifth level without a word. It does not host or proxy the model, send anything to a server, or estimate what a session cost.

Under the hood

React, Vite and Tailwind on an Electron main process; node-pty with xterm.js for the terminals. The landing shows the app’s own panels at true scale, each running its own small state machine, over a page written in my handwriting.

Where it stands

Free and open source, used every day on this site and others. The download is on the landing.