Orca is an open-source ADE ("Agent Development Environment") for running several AI coding agents in parallel. Instead of one agent working on one checkout, every task gets its own git worktree, its own agent terminal, and its own browser tab — so you can fan a single prompt across five agents, compare the diffs, and merge the winner.

It is a desktop app (Electron + TypeScript + React) built by Stably AI, MIT-licensed, with companion apps for iOS and Android. It does not provide models or a subscription of its own: you bring agents you already have (Claude Code, Codex, opencode, Pi, crush, …) and Orca orchestrates them locally.

Direct builds for macOS (ARM64/Intel), Windows (.exe) and Linux (AppImage) are on the releases page. orca serve runs it headless on a Linux server. The mobile companion pairs with the desktop app to monitor agents and send follow-ups from a phone.

The worktree is Orca's unit of work. Creating one asks for a task name and a start-from ref (base branch, local branch, commit SHA, or remote branch), and optionally links a GitHub/GitLab/Linear/Jira issue. Creation is async — the dialog closes while git fetch and git worktree add run in the background.

Any CLI agent works — if it runs in a terminal, it runs in Orca. 40+ are preconfigured, including Claude Code, Codex, Grok, Cursor CLI, GitHub Copilot, opencode, Pi, Amp, Cline, Continue, Devin, Goose, Droid, Kimi, Qwen Code, Mistral Vibe and Rovo Dev. Custom agents can be added by pointing Orca at their command ("bring your own agent").

Monaco (VS Code) editor with autosave everywhere, native file explorer, Quick Open / Jump Palette across worktrees, files, agents and commands, and rich previews for Markdown, Mermaid, PDFs and images. Files and images can be dragged directly into an agent prompt.

Terminals are Ghostty-class: WebGL rendering, infinite splits, scrollback that survives restarts.

Each worktree gets its own embedded Chromium tab. Design Mode lets you click any element in that real browser window and send its HTML, CSS and a cropped screenshot into the agent's prompt. There is also an embedded iOS Simulator agents can drive.

SSH worktrees run the agent on a remote box with full file editing, git and terminals, including auto-reconnect and port forwarding.

Agents can drive Orca itself, which is what makes multi-agent workflows scriptable. Most commands take --json and a --worktree selector.

orca terminal send + wait + read is the primitive behind agent-to-agent handoff: one agent starts another in a fresh worktree and blocks on its result.

The orca computer subcommand lets agents operate native desktop apps outside the embedded browser: list-apps and get-app-state (accessibility tree) to look, then click, set-value, type-text, press-key, hotkey, paste-text, scroll, drag, perform-secondary-action to act. orca computer capabilities and orca computer permissions report what is available and what is still ungranted.

The pattern is snapshot → act → snapshot: element indexes are scoped to the latest get-app-state and may be sparse, so state must be re-read after every action. Prefer semantic actions (click, set-value) over raw keystrokes — they target accessibility elements and survive UI changes. Requires Accessibility permission, plus Screen Recording on macOS, granted to Orca Computer Use in System Settings.