Orca ADE: Run Claude Code & Codex in Parallel Worktrees

Orca is a free, open-source desktop app, licensed under MIT, that runs multiple CLI coding agents such as Claude Code and Codex in parallel, each isolated in its own git worktree. It's built by Stably AI, a San Francisco startup out of Y Combinator, and describes itself as an "Agent Development Environment" (ADE) that bundles terminals, diffs, a browser, and a CLI into one app.

As of this writing (September 20, 2026), the latest release is v1.4.205, shipped on 2026-09-17. The GitHub repository stablyai/orca has around 73,000 stars and 4,800 forks, and the app is built with Electron and TypeScript. This article walks through Orca's main features, installation, basic usage, and how it compares to existing tools, based only on the official site, official docs, and GitHub.

The basic flow is simple. You register an existing repository once, then create a task-scoped unit of work — a worktree — for it. Each worktree bundles a real git worktree, a dedicated branch, an agent terminal, and a dedicated browser tab. After an agent works on a task, you review the changes in the diff viewer, leave annotations on lines you care about and send them back to the agent, then commit, push, and open a PR, all from inside Orca. Because worktrees are isolated, work on one task never touches another task's files or branch.

Design Mode translates visual feedback about the UI into code-level context automatically. Toggle Design Mode from the toolbar of the real Chromium window that runs alongside each worktree, then click the element you want changed. Its outer HTML, surrounding context, computed styles, a cropped screenshot, and — if source maps are available — the source file and line number are all handed to the agent's terminal as a single attachment. You can queue up multiple annotations in a tray before sending them together. Once the agent fixes the code, the page hot-reloads so you can click the same element again to verify the result.

Annotate AI Diff turns code-review comments directly into agent instructions. Hovering a line in the diff viewer reveals a + button (or press c) to leave a Markdown comment. Clicking "Send to agent" at the top bundles every comment, with line numbers, into a single prompt that can go to an existing agent session or start a new one. Comments track their line even as edits shift things around, and resolved comments can be collapsed with "Resolve." Note that the "Send Review Notes to Agent" shortcut has no default key binding — you need to set one under Settings → Shortcuts.

Outside of those two options, Orca is downloaded directly from GitHub Releases or the official site: a .dmg for macOS (separate Apple Silicon and Intel builds), a .exe for Windows, and an AppImage (self-updating), .deb, or .rpm for Linux (updated via the package manager). The app itself is free and open source under MIT, but Orca only orchestrates agent execution — you bring your own subscription for each agent. Prompts and code go directly to whichever provider you've configured, not through Orca. On first launch it asks for access to your home directory and can optionally import existing terminal settings or API credentials. On Windows, pick PowerShell or CMD as the default shell under Settings → Terminal. Updates automatically follow the stable channel (Shift-click the update check to get an RC build).

- Click "Add Repo" in the sidebar and pick a local checkout; its default branch becomes the base ref
- Click the "+" next to the repo name, name the task (e.g. fix-login-race), and create a worktree. It can start from the base ref (origin/main), or any other branch or commit
- Pick an agent (Claude Code, Codex, Cursor CLI, etc.) from the terminal dropdown
- Repeat steps 2–3 twice more to create three worktrees and send them the same prompt
- Drag the tabs to the side or bottom to split the view and watch all three run at once
- Compare the results in the diff viewer, use Annotate AI Diff to send follow-up instructions to the best candidate, then commit and push from Orca. Delete the worktrees you didn't pick (their branches are deleted automatically)

A fresh worktree is a clean checkout, so it has no installed dependencies like node_modules. The official docs list three ways to deal with this.

Branch names are generated automatically from the workspace name or a linked issue (emoji get converted to shortcodes like rocket), or you can set one explicitly under Advanced. Deleting a worktree removes its directory and branch after confirmation; branches with unmerged commits are kept and surfaced in the review screen instead. Worktrees created outside Orca with git worktree add don't show up in the sidebar until you explicitly reveal them. More worktrees mean more file-watching and memory use, so official troubleshooting guidance recommends closing worktrees you're not using — splits with many browser tabs are especially heavy on RAM.

Orca offers four ways to run, differing in where the UI and the agent runtime actually live: Local, which stays entirely on your laptop; SSH, which connects to an existing remote host; Remote Orca Server, a dedicated always-on machine shared by multiple clients; and the experimental Cloud VM, which spins up a disposable environment per task.

To run over SSH, add a host and test the connection under Settings → SSH (the remote side needs git), then pick that host under "Run on" when creating a worktree; the agent stops if the SSH connection drops. Remote Orca Server requires installing both Orca and Tailscale on the server and the client: on the server, toggle "Advertise this app as a server" under Settings → Remote Orca Servers and issue a New Link, then paste that link into Add Server on the client. A headless Linux server starts with the command above. Cloud VM is an experimental feature enabled under Settings → Experimental; you install the Cloud VM skill and have the agent write environment recipes as environmentRecipes in orca.yaml. Example supported providers include Vercel Sandbox, Fly, Modal, a plain SSH host, and local Docker. Orca does not sell managed hosting itself — usage fees go directly to whichever provider you choose.

Orca also exposes a CLI, the orca binary, registered from Settings → General → Orca CLI (macOS creates a shim in ~/.local/bin). Nearly every command accepts --json, and the CLI is designed as much for agents to drive Orca as for humans. To let an agent use the Orca CLI, run npx skills add https://github.com/stablyai/orca --skill orca-cli (or orca skills install --skill orca-cli in a headless environment).

Enabling Settings → Experimental unlocks an orchestration layer for coordinating multiple agents, built around Runs (a namespace plus a coordinator inbox), Tasks, Dispatches, workers, messages, and decision gates.

A worker sends orca orchestration send --type worker_done ... --task-id <taskId> --dispatch-id <dispatchId> --outcome succeeded exactly once when a task finishes, whether it succeeded or failed. For a one-off instruction, the docs note that orca terminal send alone is enough — orchestration isn't required. For recurring work, Orca uses Automations.

Triggers can be hourly, daily, weekdays, or weekly, or a cron expression or RRULE. --timezone takes an IANA name, --precheck "command" skips a run when the check fails, --reuse-session reuses the same session, and --missed-run-grace-minutes 30 sets a grace window for missed runs. The docs recommend creating automations with --disabled first and testing before turning them on. Running an automation requires either the desktop app to be open or a Remote Orca Server; results show up on the Automations page and via orca automations runs --id <automationId> --json.

Mobile Companion pairs a phone with the desktop app using a one-time code shown in the desktop account menu. It connects either through Orca Relay (which requires sign-in, with an optional fast LAN path) or a direct LAN connection. From the phone you can see the status of every worktree (working, done, or blocked), use a chat UI or the raw terminal, get push notifications on completion, send short replies — text, photo, or voice — to a blocked task, switch accounts and check usage, run Quick Commands, review, stage, and commit changed files, and even create new worktrees. The official description frames it as "not a full editor, but a remote control for your running desktop," and the connection drops when the desktop app is closed. The docs don't specify encryption details.

For a closer look at cmux and Herdr specifically, see our cmux deep dive and our Herdr deep dive. These tools all ship fast, so always check the official docs for the current spec.

- Each agent CLI (Claude Code, Codex, etc.) still needs its own install and authentication — Orca only orchestrates them
- More worktrees mean more file-watching and memory use; close ones you're not using, and note that splits with many browser tabs are especially RAM-heavy
- Usage and cost figures are estimates read from local usage files an agent writes, not from direct API calls, so they can lag; check each provider's console for accurate billing
- Local mode stops when your laptop sleeps, so consider SSH or Remote Orca Server if you need something always on
- Telemetry (via PostHog Cloud, anonymous usage events only — no file paths, prompts, responses, or terminal content) is on by default; opt out via Settings → Privacy or the environment variables DO_NOT_TRACK=1 / ORCA_TELEMETRY_DISABLED=1
- Releases ship very fast, roughly every one to two days — teams using Orca for real work should agree on when to update

The official docs don't mention Japanese-language UI or IME support one way or the other. For the basics of Claude Code itself, see our complete Claude Code guide.

Orca runs multiple coding agents in parallel, one per worktree, and tries to fold diff review and in-browser verification into a single app. It's free and open source under MIT, and it also reaches beyond local use into remote execution over SSH or a Remote Orca Server, plus automation through its CLI and orchestration layer.

That said, each agent's own subscription and authentication is still your responsibility, and more worktrees mean more machine resources consumed. Starting locally with one or two worktrees and expanding to SSH, Remote Orca Server, or Cloud VM as needed looks like a reasonable way to adopt it.

- https://www.onorca.dev/
- https://www.onorca.dev/docs
- https://www.onorca.dev/docs/install
- https://www.onorca.dev/docs/ways-to-run
- https://www.onorca.dev/docs/cli/overview
- https://www.onorca.dev/docs/telemetry
- https://github.com/stablyai/orca