Orca is a desktop and mobile workspace for running a fleet of coding agents in parallel, each in its own git worktree. It is aimed at developers who already pay for more than one coding agent — Codex, Claude Code, Cursor's agent, OpenCode, Pi — and who have got tired of juggling them in separate terminal windows and hand-made branches. The project calls itself an ADE, an agent development environment: instead of being the agent, it is the place where several agents work and where you watch them. It is published by Stably under the MIT license, written in TypeScript, and has drawn about 82,000 stars and 5,300 forks since the repository appeared in March 2026.
What it does
Orca runs any supported coding agent on your own subscription, so you keep the plan you already have rather than buying tokens through a middleman. From one place you can:
- start several agents side by side on the same task or on different tasks
- see each one's progress, tracked in a single view rather than across scattered terminals
- steer an agent from a phone companion app, including notifications when one finishes
- push the heavy work to a remote machine over SSH instead of running it on your laptop
The desktop app is published for macOS, Windows and Linux. The mobile side is a companion to it, not a separate product: the point is that the same set of running agents is reachable from either, and from a remote runtime, in sync.
How it works
The mechanism the whole tool is built around is the git worktree. Each agent gets its own worktree and its own branch, so two agents editing the same file never collide in one checkout, and nothing needs stashing between runs. One prompt can fan out to five agents on five branches at the same time.
That changes how you review. Rather than accepting or rejecting a single agent's patch, you let the worktrees finish, look at what each produced, and keep the best one. The others are ordinary branches you can discard. Because the unit of work is a branch rather than a live editor buffer, the same run can be inspected from the desktop app, from the phone, or on the remote box that actually executed it.
Getting started
The download page at onorca.dev/download carries prebuilt desktop builds for the three platforms, which is the intended route in — you do not have to build the TypeScript source to use it. After that the setup you supply yourself is the agents: Orca drives the CLIs and subscriptions you already hold, so whichever of Codex, Claude Code, Cursor, OpenCode or Pi you have access to is what you can run in it. A git repository is the other prerequisite in practice, since worktrees are how tasks are isolated.
The project keeps a Discord and an X account for questions and release news, and the README is translated into several languages. Commits were still landing at the end of September 2026, so the release cadence is current rather than historical.
When to use it / when not
Orca earns its place when you genuinely run more than one agent, or one agent more than once. Racing three attempts at the same refactor and picking the winner is the case it was designed for, and so is the half-attended workflow where you kick off a long job, walk away, and get a phone notification when it lands. The remote runtime matters if your machine is a laptop and the job is not laptop-sized.
It is overhead you do not need if your habit is one agent in one terminal on one branch. The same is true if you have access to only one agent: much of Orca's value is the side-by-side comparison, and with a single subscription that collapses to running the agent's own interface with extra steps. And if you dislike branch-heavy workflows, the worktree model is not optional here — it is the design.
Alternatives
The honest comparison is the manual version of the same thing: each agent in its own terminal tab, worktrees created by hand, branches tracked in your head. That works, and it costs nothing new to learn; the script behind the video makes the case that it also costs hours. Each agent also ships its own front end, which is the simpler choice when one agent is all you use. The materials here do not compare Orca against other agent orchestrators, so treat the field beyond that as unmeasured.
Anyone who has started paying for two or three coding agents should look at Orca seriously, because it addresses the specific waste that comes with that situation: attention split across windows, checkouts fighting each other, and no easy way to compare two attempts at the same change. Its star count says the problem is widely felt. The project is young, so expect rough edges, but the core idea — one agent, one worktree, one branch, all of them visible at once — is a sound answer to a real and recent problem.