stablyai/orca

Orca is a desktop orchestration shell for CLI-based coding agents — the kind that run in a terminal, like Claude Code, Codex, or Goose. Its central idea is the parallel worktree: you fan a single prompt out to multiple agents simultaneously, each working in an isolated git worktree, and then compare the results before merging whichever one you prefer. That is a genuinely useful primitive. The cost of running five agents at once on the same task is mostly compute; the cost of context-switching between five separate terminal sessions to compare their outputs is human attention. Orca collapses the second cost.

The feature set around that core is dense. There is a WebGL-accelerated terminal with infinite splits and scrollback that persists across restarts. A Design Mode embeds a real Chromium window and lets you click any UI element to pipe its HTML, CSS, and a cropped screenshot directly into an agent prompt — useful for front-end iteration loops where the agent needs to see what it is actually producing. Diff annotation lets you drop inline comments on any changed line and route them back to the agent without leaving the app. SSH worktrees extend all of this to remote machines, with auto-reconnect and port forwarding. There is also built-in usage tracking that surfaces Claude and Codex consumption and rate-limit resets, plus hot-swap account switching without re-logging in.

The mobile companion is the most unusual piece. It pairs with the desktop app so you can monitor running agents from your phone, get notified when one finishes, and send follow-up instructions from wherever you are. The Android build is distributed as a direct APK download rather than through the Play Store, which is worth knowing before you install it.

Agent support is deliberately broad — the README lists well over two dozen named CLI agents and ends with a catch-all for anything that runs in a terminal. That breadth is the right call architecturally; Orca is not betting on any one agent winning.

One thing the README does not address: there is no discussion of how context is managed across parallel worktrees when agents diverge significantly. The privacy documentation is also offloaded to a separate page rather than summarized inline — the telemetry opt-out exists, but you have to go find it.

The project ships under the MIT license, runs on macOS, Windows, and Linux, and the team describes a daily shipping cadence — the changelog is explicitly called the real feature list because the README cannot keep up. That is either a sign of genuine momentum or a warning that the surface area is growing faster than the documentation. Probably both.