Orca (stablyai/orca) – ADE for orchestrating a fleet of parallel AI agents

Orca is an open ADE (Agent Development Environment) from the Stably team. The idea is simple and powerful: run multiple CLI coding agents in parallel, each in its own isolated git worktree, and manage the entire fleet from one place. The AI Orchestrator for 100x builders.

If it runs in a terminal, it runs in Orca, if it runs in a terminal, it runs in Orca. You use your own subscriptions to agents, Orca only conducts them.

License - MIT, the project is backed by Y Combinator, website - onOrca.dev. There are about 57,000 stars on GitHub.

Orca is not another agent, but a ** orchestrator** on top of the existing ones: Claude Code, Codex, Grok, Cursor, Copilot, OpenCode, Pi, Goose, Cline, Kiro and many others (25+).

## In Short: What Orca Does

• Runs any CLI agent** side-by-side, each in a separate git worktree.
• Spreads one prompt on several agents at once, then compare the results and destroy the best.
• Gives a *mobile companion to monitor agents from the phone.
• It works on desktop (macOS, Windows, Linux), mobile (iOS, Android) and VPS (headless mode).
• Uses your subscriptions - does not take money for tokens.

## Key idea: parallel worktree

It's Orca's heart. Instead of running one agent in one terminal, you:

1. Write a task.
2. Distribute it to multiple agents at once (e.g. Claude Code + Codex + Grok).
3. Everyone works in their own insulated git worktree – the common branch doesn’t get dirty.
4. Ready to compare the diffs and know the winner.

This approach removes the main pain of multi-agent work – the chaos of dozens of terminals and branches, where it is easy to lose who did what.

## Key opportunities

• Parallel worktree. A fan from one prompt over several agents in isolated git worktree, then a comparison and a merz of the best.
• Orca Mobile. Monitoring and managing agents from the phone, termination notifications and sending follow-up remotely. QR pairing in a few seconds.
• Ghostty level terminals. WebGL rendering, endless splits, scrollback that survives restart.
• Design Mode. Click on the UI element in the real Chromium window and its HTML, CSS and cropped screenshot are sent to the prompt agent.
• Native integration with GitHub and Linear. View PR, issue and project boards directly in the application; open worktree from the task.
• SSH worktree. Running agents on a remote machine with file editing, git, terminals, auto-reconnect and port forwarding.
• Diff annotation. Comments on the lines of the diff are sent back to the agent - review, edit and commit without leaving Orca.
• Workspace Board. Kanban board (Todo / In Review / Done), which extinguishes the chaos of tabs and protects the attention of the person.
• Orca CLI. Scripting workshopflow teams orca worktree create, snapshot, click, fill, etc.
• Other: Drag and drop files and pictures into a prompt, VS Code-based editor, quick search, switching accounts with limit tracking, rich repository previews (Markdown, images, PDF), computer use, notifications, saved commands (buttons for repetitive PR/commit-prompts) and session history.

## Supported agents

Works with any CLI agent**, among them: Claude Code, Codex, Grok, Cursor, GitHub Copilot, OpenCode, Pi, Goose, Cline, Continue, Kiro and many others (25+).

## Technology

• TypeScript + Electron + Vite (electron.vite.config.ts, vite.web.config.ts).
• pnpm workspaces, ling/formatting via oxlint/oxfmt, git-hooks on Husky.
• Components: desktop (native), mobile, web; system of plugins and “skills”.

## Installation

# Homebrew (macOS)
brew install --cask stablyai/orca/orca

# Arch Linux (AUR)
yay -S stably-orca-bin

# Headless-режим на сервере / VPS
orca serve

• Desktop (macOS, Windows, Linux): download from onOrca.dev or finished builds.
• Mobile: iOS - App Store / TestFlight, Android - APK.

е️ Important: On the organization's Stably page, assemblies are marked as "Unvetted: no tests, signed but notarized." Not for general use – that is, without tests, signed but notarized, not for everyday use. The project is actively developing, versions come out often (at the time of reviews - v1.4.x).

### How much does it cost

Orca itself is free and open (MIT). It doesn’t charge for tokens — you plug in your own subscriptions to agents (Claude Code, Codex, etc.) and pay directly to their providers. There are usage tracking and limits on accounts.

## Orca is better

• ** Removes the chaos of multi-agent work.** Instead of scattering terminals, tabs and branches - a single panel, Kanban board and status icons. Reviewers note that this is what “strongly suppresses the disorder” and preserves attention.
• It works with any CLI agent and your subscriptions – you can mix Claude Code, Codex, Grok in one task.
• Isolation through worktree. Agents do not spoil the common branch, diffs are compared safely.
• Mobility. You can actually monitor agents from your phone and send instructions, which is a rare feature.
• Design Mode and computer use. Click on element → context to prompt speeds up front-end tasks.
• Works on VPS. Headless-mode orca serve for remote machines.

## Orca is worse

• **No use for a single script. ** If you have enough **one * Claude Code in one terminal - the benefit is minimal, the overhead does not pay off.
• Unvetted assemblies. Officially marked as not for general use: no tests, not notarized. It's about early adulthood.
• Input curve. Computer Use and Orca CLI must be manually enabled in the settings, they are not for everyone at once.
• Orca does not improve the quality of the edits themselves - it only organizes agents. The mind is still in connected models.
• Electron application. It is heavier than purely terminal resources.
• No branching of the context of the session (fork of the existing dialogue) – stated as a requested but missing feature.

## What the reviews say

• masa wunder (note.com, detailed analysis): S+/SS-class rating. Most praised is isolation through worktree, status review left menu, Workspace Board (Kanban), saved commands with AI-generated commit messages, session history and instant QR pairing with your phone. The main conclusion: Orca goes “beyond ostentatious automation”, keeping the screen in order – it is clear where to look and what to review after launching several AIs.
• For those who are happy with one Claude Code in one terminal, there is almost no profit.
• Positioning: Several reviews (including the Japanese parsing on note.com) agree that Orca relieves the fatigue of juggling Ghostty + VS Code + agent desktop applications by putting everything together in one environment.

## What do they compare

The key difference is that Orca solves the “many agents at once” problem, whereas oh-my-pi or Claude Code is one very smart agent. They can even be used together: Orca as a conductor, these agents as performers.

## Who does and who doesn't

**This is for those who parallel multiple agents (Claude Code + Codex + ...) and want to compare results; who reads diffs, manages commits and PR; who needs control from a phone or run on a VPS; teams with many tasks at the same time.

Not suitable: for those who have enough agent in one terminal; who do not want a heavy Electron application; who are waiting for a "stable release for production" - builds are labeled as unvetted.

## Outcome

Orca is a conductor for a fleet of AI agents. It does not make the edits smarter, but solves the real pain of those who chase multiple agents at once: isolation through git worktree, single status panel, Kanban board, mobile control and launch on VPS. For a 100x developer who parallels the task and revisits the results himself, this is a godsend.

The flip side is early maturity (non-public builds), Electron weight, and the complete uselessness of the script if one agent is enough. If you’re already juggling Claude Code, Codex, and a pair of terminals, Orca is worth a try.

Sources: репозиторий stablyai/orca, onOrca.dev, страница организации Stably, обзор masa_wunder на note.com, разбор кудокуры на note.com. *