How Do I Get Started with Orca? From Install to Your First Parallel Task

Last updated: 2026-10-25

Orca is a free, open-source desktop app from Stably AI that calls itself an ADE, an Agent Development Environment. It puts command-line coding agents such as Claude Code, Codex, OpenCode, and Pi into one window, giving each agent its own isolated git worktree so several agents can work on the same task in parallel. This guide walks through that workflow once, from installation and connecting your existing subscriptions to opening your first worktree, fanning one prompt out to several agents, comparing their diffs, and merging the winner.

Where do I download and install Orca?

Start at the front door. The official site is onorca.dev, and installers are also available from the GitHub repository releases page. On macOS, the shortest path is Homebrew: brew install --cask stablyai/orca/orca. On Arch Linux, use the AUR package: yay -S stably-orca-bin. Windows users install the .exe directly, and Linux users grab the AppImage. Desktop builds cover macOS, Windows, and Linux, so once the installer finishes you can open the app and land on the main screen without setting up a runtime first.

How do I connect the agent subscriptions I already pay for?

Orca does not sell models, and it does not handle model billing. Every cent of agent cost runs through your own subscriptions. After opening Orca, sign in with the agent accounts you already have, such as Claude Code or Codex, with no need to register a new subscription. A built-in account switcher shows usage and rate-limit reset times for Claude and Codex, and lets you hop between accounts without logging out and back in, so hitting a limit on one agent does not stop the whole session. Once connected, those agents can work directly inside the Orca window.

How do I open my first worktree?

Open a git repository first, then create a worktree for the task you want handled. A worktree is the basic unit of work in Orca: every agent works inside its own isolated worktree, so agents never step on each other's files, which removes the whole class of conflicts that comes from several processes sharing one working directory. A Chinese summary of the Orca README, published by Tenacme, uses the keywords parallel agents, agent IDE, and multi-agent scheduling, and it is worth reading alongside the original. One task gets one worktree, and worktrees do not leak into each other. That isolation is the precondition for everything that follows.

How do I fan one prompt out to several agents?

With worktrees ready, type one prompt, pick the agents you want to run in parallel, and Orca starts them at the same time in separate worktrees. You can point them all at the same task with the same prompt and let them race, or you can hand different worktrees completely different jobs, even with different agents attached, for example one agent on the backend and another on the frontend. A hands-on guide on Bilibili sums the flow up as asking once, letting several agents work at the same time, and merging the best result, and following it once makes the pattern click. While the agents run, you do not need to watch every terminal: Orca collects each agent's status into one view.

How do I compare diffs and merge the winner?

When the agents finish, go back to the Orca window and compare their diffs side by side. Read line by line, and when a line bothers you, write an annotation right on that line and send it back to the agent for another pass. Once you have seen enough, merge the result you like and discard the rest. Review, revision, and submission close the loop inside one window, with no need to jump between apps. Comparison is the heart of Orca: several answers laid out together, with the human keeping the job of choosing and the machine doing the job of preparing options.

How do I follow up from my phone or a remote machine?

You do not have to stay at your desk. Companion apps for iOS and Android let you check agent status remotely, get a push notification when a task finishes, and send follow-up instructions from anywhere. Heavy jobs can move to a beefier remote server as SSH remote worktrees, with automatic reconnection and port forwarding included. Design Mode in the built-in browser lets you click any element in a real Chromium window and drop its HTML, CSS, and a cropped screenshot straight into the prompt. Orca also ships its own command line, with commands such as orca worktree create, snapshot, click, and fill, which make the whole workflow scriptable. Treat these four as a first taste, and dig deeper once the basics feel natural.

What details are worth knowing on the first run?

Pick the method that matches your machine. Homebrew is the smoothest path on macOS, and it keeps upgrades easy later. Direct installers suit Windows and Linux just as well. Mobile apps are a separate step: iOS from the App Store or TestFlight, Android as a direct APK download, paired with the desktop app afterwards. Get the desktop app running first and add mobile later, since there is no need to install everything at once.

Subscriptions stay separate. Each agent bills against its own usage, and Orca only gathers them for scheduling without merging the bills. When one agent hits a limit, switching accounts in the switcher keeps the parallel task moving instead of forcing a wait.

A worktree usually maps to one branch: a separate working directory for that branch, with changes kept apart and history shared. Once a task is merged, the remaining worktrees can be dropped instead of kept around. Keeping one worktree per task stays tidy over time.

Prompts can go out two ways. One is racing: the same prompt to several agents on the same task, which fits moments when you are unsure how to approach it and want to see several approaches first. The other is splitting: different worktrees doing different jobs, possibly with different agents attached, which fits tasks that divide naturally. After dispatch, there is no need to watch every terminal, since each agent's progress and input requests show up in one view.

Annotations can go back and forth. When a first round misses the mark, write the objection on the diff line and send it back, then compare again until one result is worth merging. Set your own acceptance bar before merging: whether the change does what you asked and whether the logic holds together. The merged result joins the main line and the rest are discarded, which closes the task.

These pieces combine well. Dispatch from the desk during the day, check progress from the phone in the evening, get a push when a task finishes, and send a follow-up from the phone. Move heavy jobs to a remote server and keep only a light control end locally. For a visual front-end change, Design Mode turns a click on a page element into HTML, CSS, and a screenshot inside the prompt, with no manual description of the interface. Repeated flows can become Orca CLI scripts, so the next run takes a single command.

Different tasks reward different agent combinations. Some tasks suit racing, others suit splitting, and the only way to learn which combination fits your work is to run the loop again with a different task and a different mix. Once this loop runs end to end, the rest is repetition: open a worktree, dispatch, compare, merge, with phone follow-up and remote worktrees added when useful.

Should non-programmers start with Orca?

Honestly, Orca assumes you already use coding agents and already pay their subscriptions, and it assumes you are comfortable with git worktrees, terminal work, and diff review, because the whole acceptance step is reading diffs, writing annotations, and merging. A product built in the opposite direction is aiglade: 「Aiglade 是一个 Windows 桌面软件，把 Claude Code、Codex CLI、Hermes、OpenClaw 等多个 AI 装进一个中文窗口，带中文向导，装上就能用，不用会编程。」Its slogan is 「不用先学会技术，也能让 AI 替你干活」. In one line: people who code and want to direct several coding agents at once should look at Orca, and people who do not code and just want AI to do the work should look at aiglade.

References