Zenolab.app
Once you’re comfortable with one coding agent, the next question is how to run several at once without them editing the same files. Claude Code Desktop answers that for Claude sessions, as covered in our parallel sessions guide. Orca answers it for every agent at once: Claude Code, Codex, Pi, Oh My Pi, and dozens more, side by side in one app.
Orca is on version 1.4.235 as of October 20, 2026, and the team says it ships daily. This guide covers what Orca is, how to install it, the core workflow, and when it’s worth using. It is based on the Orca README and the Orca docs.
A note on the name: several unrelated products are called Orca, including the GNOME Orca screen reader. This post is about the coding tool at onorca.dev and stablyai/orca.
What Orca is
Orca calls itself an agent development environment, or ADE. In the docs’ words, it is “a desktop IDE for running multiple AI coding agents side by side. Every task gets its own git worktree, its own agent terminal, and its own browser tab.”
The docs are just as clear about what Orca is not:
- Not a model. You bring the agents and subscriptions you already have.
- Not a git replacement. Every worktree is a real git worktree, and you can
cdinto it and use plain git. - Not a hosted VPS product. Orca runs on your desktop by default. Remote modes use machines and cloud accounts you control.
Key facts:
- Free and open source under the MIT license
- Built by Stably AI
- Desktop app for macOS, Windows, and Linux, plus a mobile companion for iOS and Android (in beta)
How to install Orca
Download it from onorca.dev/download or GitHub Releases. Builds exist for macOS (Apple Silicon and Intel), Windows, and Linux (AppImage, .deb, and .rpm).
With a package manager:
# macOS (Homebrew)
brew install --cask stablyai/orca/orca
# Arch Linux (AUR)
yay -S stably-orca-bin
On first launch, Orca asks for access to your home directory so it can add repos, and offers to import ~/.claude, ~/.codex, and Ghostty terminal settings if it finds them. Orca updates itself on the stable channel by default.
You still install each agent yourself. For Claude Code, install it as usual (see our getting started guide) and log in once from any terminal. Orca picks up ~/.claude automatically.
The core workflow: one task, one worktree
The first session guide sums up Orca in one line: add a repo, create a worktree, start an agent, split panes, review the diff, ship. Here is that flow:
- Add a repo. Click Add Repo and point Orca at a local checkout. Your default branch becomes the base ref, usually
origin/main. - Create a worktree. Click the
+next to the repo and name the task, for examplefix-login-race. Orca runsgit worktree add, so the task gets its own branch and its own files on disk. - Pick an agent. A terminal opens with an agent picker. Choose Claude Code, Codex, Cursor CLI, or another agent. Orca launches it in that worktree.
- Watch them work. Drag tabs to the edge of a pane to split the view and watch several agents at once.
- Review and ship. Open the diff view, leave comments, then commit, push, and open a PR from inside Orca.
- Clean up. Deleting a worktree removes its folder and branch in one click, after a confirmation.
Because each agent has its own checkout, they never step on each other’s files. That is the same reason Claude Code uses worktrees for isolation, explained in our isolation deep dive.
Race three agents on the same task
The docs call this “Orca’s killer move.” Create three worktrees from the same ref, start a different agent in each, and paste the same prompt into all three:
fix-bugwith Claude Codefix-bug-2with Codexfix-bug-3with Cursor CLI
When they finish, compare the three diffs, keep the best one, and delete the other two worktrees. The docs’ reasoning: different agents make different mistakes. Where all three agree, the answer is probably right. Where they disagree, you’ve found the hard part of the task.
This costs three times the tokens, so check your limits first. Orca helps there too (see usage tracking below), and our usage limits guide covers how Claude’s limits work.
Features worth knowing
Annotate AI Diff
Hover a line in the diff, click the +, and type a comment. Comments pin to the line and follow it if the diff shifts. When you’re done, send all of them back to the agent as one batch. It is a lighter loop than a full PR review, and pairs well with tools like Claude Code’s /code-review.
Attribution
Orca records which lines an agent wrote and marks them in the diff gutter. If you edit an AI-written line by hand, it flips back to human. Attribution stays inside Orca and is not committed to git.
Embedded browser and Design Mode
Each worktree can have its own Chromium tab. Turn on Design Mode, click any element on the page, and Orca sends its HTML, computed CSS, a cropped screenshot, and the source location (when a dev source map is available) to the agent. Then you type what you want changed.
Usage and rate-limit tracking
The status bar shows current usage and time to reset for Claude Code, Codex, Gemini, OpenCode, Cursor, and several other providers, with a warning when you pass 80% of a limit. You can also hot-swap between several Claude or Codex accounts without logging in again.
Remote and SSH worktrees
Point Orca at an SSH host and create a worktree there. The agent and git run on the remote machine while the editor and diff stay on your laptop. If your laptop sleeps, the agent keeps running and Orca reattaches when you reconnect. For an always-on setup, a Remote Orca Server keeps projects and agent sessions on another machine, such as a Mac mini, which you connect to over a private network like Tailscale.
The Orca CLI
The orca command lets scripts, and agents themselves, drive the app:
orca status --json
orca worktree create --repo id:<repoId> --name my-task --issue 123 --json
orca terminal send --text "continue" --enter --json
It ships with the desktop app; register it in Orca’s settings. On Linux the CLI is named orca-ide so it doesn’t clash with the GNOME screen reader.
Mobile companion
Pair the iOS or Android app with your desktop to see each worktree’s agent status, read recent terminal output, and send a quick reply when an agent is waiting for input. The mobile app is in beta.
Your existing hooks and memory still work
Orca reads each repo’s .claude/ and .codex/ config, so hooks you already have run when Orca starts the agent. CLAUDE.md and AGENTS.md are left as they are. You can also set worktree setup hooks, such as pnpm install or restoring .env files, that run after each new worktree is created.
Orca vs Claude Code Desktop
| Orca | Claude Code Desktop | |
|---|---|---|
| Agents | Claude Code, Codex, Pi, oh-my-pi, and any CLI agent | Claude |
| Isolation | Every task in its own git worktree | Optional worktree per session |
| Remote | SSH worktrees, Remote Orca Servers | Local, SSH, and WSL sessions |
| License | MIT, open source | Proprietary |
Stay with Claude Code Desktop if you only use Claude and want Anthropic’s own app.
Try Orca if you pay for more than one agent, want to race them on the same task, or want one window for every agent you run. It also fits the Pi family well: you can run Pi or Oh My Pi in one pane and Claude Code in the next.
Privacy
Packaged builds send anonymous usage events keyed by a random local ID. According to the telemetry docs, Orca never sends file contents, prompts, agent output, terminal output, repo or branch names, URLs, paths, or commit messages. To turn telemetry off, set DO_NOT_TRACK=1 or ORCA_TELEMETRY_DISABLED=1.
Orca doesn’t change what your agents are allowed to do. Each agent keeps its own permission settings, so the advice in our auto mode guide still applies inside Orca.
Sources
- stablyai/orca on GitHub (README and releases)
- Orca docs: What is Orca?, Install, Your first 3-agent session, Ways to run Orca, Orca CLI
- onorca.dev, the official site