Orca: AI Agent Orchestrator for Parallel Coding
stablyai/orca
Orca is a desktop app that puts several coding agents to work at once, each isolated in its own git worktree so their changes never collide. Reach for it if you already run Claude Code or Codex from a terminal and want to fan a prompt across multiple agents to compare outputs; skip it if you only run one agent at a time, since the worktree/orchestration layer adds nothing there.
Understanding Orca's Role in AI Development
Orca is a desktop, mobile, and VPS application built around one idea: run several coding agents β Claude Code, Codex, OpenCode, Pi, and others β side by side instead of babysitting one terminal tab at a time. The GitHub description calls it an ADE, built for a fleet of parallel agents rather than a single one. Each agent gets its own isolated git worktree, and Orca tracks all of them in one place.
Key Capabilities for Agent Orchestration
- βParallel Worktrees β fan one prompt across five agents at once, each running in an isolated git worktree, then compare the results and merge the winner.
- βAs a Ghostty-class terminal, Terminal Splits offers WebGL rendering, infinite splits, and scrollback which persists across restarts.
- βWithin Design Mode, a UI element can be clicked when it is inside an embedded Chromium window, and directly transmit its HTML, CSS, along with a cropped screenshot to an agent's prompt.
- βSSH Worktrees β run agents on a remote box with full file editing, git, and terminal access, with auto-reconnect and port forwarding.
- βGitHub & Linear integration β browse PRs, issues, and project boards inside Orca, and open a worktree straight from a task.
- βMobile Companion β iOS and Android apps pair with the desktop app to monitor agents and send follow-up prompts remotely, with a push notification when an agent finishes.
- βOrca CLI β script workflows with commands like `orca worktree create`, `snapshot`, `click`, and `fill`, so agents can drive Orca itself.
- βAnnotate AI Diffs β leave line comments on a diff and send them back to the agent for revision without leaving the app.
Who Benefits from Using Orca?
Orca is built for developers who already run coding agents like Claude Code or Codex from a terminal and want to run more than one at a time without hand-rolling a git worktree for each. It fits someone managing a fleet of agent runs across a team, or a solo builder juggling several long tasks β less so someone who only ever runs a single agent inline in their editor, since the multi-agent, multi-worktree machinery buys them nothing extra there.
Getting Started with Orca
Desktop builds for macOS, Windows, and Linux are on onOrca.dev/download, or you can grab a platform build straight from GitHub Releases (a .dmg for Apple Silicon or Intel Macs, a Windows .exe, or a Linux AppImage). Homebrew users on macOS can run `brew install --cask stablyai/orca/orca`; Arch Linux has an AUR package, `yay -S stably-orca-bin`, or `stably-orca-git` to build from source. Mobile companion apps are on the iOS App Store (or TestFlight for the beta) and as a direct Android APK download (version 0.0.37 per the README). For a headless Linux server, the README points to a separate headless server guide built around the `orca serve` command.
Strengths
- βWorks with basically any CLI-based agent, not just one vendor's β the Supported Agents list spans Claude Code and Codex to Devin, Goose, and Kiro, plus "+ any CLI agent."
- βGit worktrees are created and tracked for you, so running five agents against the same repo doesn't turn into five manually managed branches.
- βSSH Worktrees mean an agent run isn't capped by your laptop's CPU or battery β hand it to a bigger remote box and keep working locally.
- βMIT license, so there's no copyleft concern about folding it into a commercial workflow.
Current Considerations for Orca Users
- β³Orca depends on you already having the agent CLIs and their own subscriptions installed β it orchestrates Claude Code, Codex, and similar tools but doesn't include or replace any of them.
- β³The README calls its own feature list 'perpetually behind' because the team ships daily, so check the release changelog rather than the README for exactly what's current.
- β³Design Mode needs a Chromium window that Orca controls, which is one more moving part than a plain terminal-only agent runner.
- β³Android installs via a direct APK download rather than a Play Store listing, so you handle sideloading and updates yourself, unlike the iOS app.
Exploring Other AI Development Environments
Common Questions About Orca
Orca has companion apps for iOS (App Store and TestFlight) and Android (a direct APK download), built to pair with the desktop app so you can monitor agents and send follow-up prompts from your phone.
Orca works with any agent that runs in a terminal, including Claude Code, Codex, Cursor, GitHub Copilot, OpenCode, Grok, Devin, Goose, Cline, and Kiro, among others listed in the README's Supported Agents section.
Orca is released under the MIT License, so it's free to use and modify with no copyleft restrictions.
Orca supports headless Linux servers through the `orca serve` command, and the README links to a dedicated headless Linux server guide for setup details.
Orca collects anonymous usage telemetry, and the README points to its privacy and telemetry docs describing what's gathered and how to opt out.
The problem it solves
Running more than one coding agent at once usually means several terminal tabs and several git worktrees you created by hand so the agents don't stomp on each other's files. Nothing pings you when one finishes. Orca replaces the manual worktree juggling with one workspace that creates the worktree, opens the terminal, and notifies your phone when an agent is done.
Best use cases
- β’Fan the same prompt to Claude Code, Codex, and OpenCode at once, then diff the three worktrees and merge whichever result is best.
- β’Kick off a long agent run on an SSH-connected VPS before leaving your desk, then check progress and approve a diff from the mobile app later.
- β’Use Design Mode to click a broken component in a running app and hand the agent the exact HTML/CSS instead of describing the bug in prose.
- β’Review a batch of agent-generated diffs by leaving inline comments and re-running the agent from the same screen, instead of round-tripping through a separate PR tool.