Orca (stablyai/orca): The Open-Source Agent Orchestrator With Worktrees, Native Chat and a Phone Companion

If you already run more than one coding agent at a time, the hard part is not the agents. It is keeping track of them: which worktree each one is in, which one finished, which one needs an answer, and which result to merge. Orca is an open-source app built around that problem, and it is trending on GitHub this week with a repo page showing about 94.8k stars and 6.1k forks. This guide covers what it does, how to install it, how it compares to alternatives we have covered, and what to check before you hand it your repos.

Everything below is from the project's README and its releases page as read on October 17, 2026, plus a Hacker News thread from the project's launch. We did not run the app ourselves, so treat feature descriptions as the project's claims.

TL;DR: questions people ask first

What Orca actually does

The README's tagline is "The AI Orchestrator for 100x builders" and its one-line pitch is to run Codex, Claude Code, OpenCode or Pi side by side, each in its own worktree, tracked in one place. The feature list in the README, in the project's words and ours:

• Parallel worktrees. Fan one prompt across five agents, each in an isolated git worktree, then compare results and merge the winner. A git worktree is a second checkout of the same repository, so agents do not overwrite each other's files.
• Terminal splits. A terminal with WebGL rendering, unlimited splits and scrollback that survives restarts.
• Mobile companion. Monitor agents from your phone, get notified when one finishes, and send follow-ups. The relay that pairs the phone with a desktop host lives in the repo under cloud/.
• Design Mode. Click a UI element in a Chromium window to send its HTML, CSS and a cropped screenshot into an agent's prompt.
• GitHub and Linear, natively. Browse pull requests, issues and boards, and open a worktree from any task. The latest release notes add Jira planning with boards and saved JQL.
• SSH worktrees. Run agents on a remote machine with file editing, git and terminals, with auto-reconnect and port forwarding.
• Annotate AI diffs. Comment on any diff line and send the comments back to the agent.
• Orca CLI. Agents can drive Orca too, with commands like orca worktree create, snapshot, click and fill.
• Account switcher and usage tracking. See Claude and Codex usage and rate-limit resets, and hot-swap accounts without logging in again.
• Computer Use. Let agents operate desktop apps and visible UI when a workflow needs it.

The README also says the team ships daily and that the changelog is the real feature list. The releases page backs that up: v1.4.227 on October 16 and v1.4.229 on October 17, plus several Android pre-releases.

What changed in the latest release

The v1.4.229 notes describe, among other things:

• Native chat (experimental): review turn-by-turn edits inline, expand diagrams, and see Claude goals and Codex context usage.
• Shared skills: installation now supports 20 additional agents.
• Phone: workspace filters, grouping and sorting follow the desktop's choices.
• SSH and remote terminals: reconnect fixes preserve split panes and paste order.
• Jira planning: issue lists and boards, with agents able to read and update tickets through the Orca CLI.

The earlier v1.4.227 notes added a per-chat permission mode for Claude and Codex native chats, a requirement of Codex 0.136.0 or later for chats, and browser automation fixes so agent actions return your keyboard. The notes also list known issues: remote-host features depend on the host version, and the orca command on SSH hosts keeps saved control preferences until you enable it in the host's advanced settings.

How to install it

Per the README:

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

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

Or download builds for macOS (Apple Silicon and Intel), Windows or Linux AppImage from the project's site. For the phone, install the iOS app from the App Store or the Android APK, then pair it with your desktop host. A first-run sanity check we would use: open a small throwaway repo, create two worktrees, start one agent in each with the same prompt, and compare before pointing it at real work.

How it compares to what we have covered

Orca sits in a crowded category of tools for managing several agents. On Hacker News, one commenter, bartwaardenburg, asked how it differs from cmux, a terminal we covered in cmux for AI coding agents, saying they were happily using cmux; the Show HN title framed Orca as an open-source "Conductor + Ghostty" (thread). Another commenter, mnogueir, said it worked well for them and was the first tool that worked well on Linux. These are individual reports, not benchmarks.

• AgentPlane is a local control plane for running several agents side by side; Orca adds a phone companion, native chat and worktree-per-agent defaults.
• OpenCode is an agent, not an orchestrator; Orca lists it as one of the agents it can host.
• Claude Code commands remain the foundation if Claude Code is your main agent; Orca wraps rather than replaces it.
• Skills packs such as ECC work inside an agent; Orca's shared-skills feature installs skills across many agents at once.

Another HN user, johntash, noted that when the project launched there was no LICENSE file in the repository even though the author said MIT; the current repo lists an MIT license file and the README states MIT, so that early issue appears resolved.

A sensible first hour with Orca

If you want to evaluate it without risking real work, this is the order we would follow, based on the README's own feature list.

1. Install the desktop app with Homebrew or a direct download and open a small repository you do not mind breaking, such as a scratch project or a fork.
2. Create two worktrees from the same branch and start a different agent in each, for example Claude Code in one and Codex in the other. Give both the same small task, such as adding a unit test or fixing a typo-level bug.
3. Watch the status view. The point of the tool is knowing which agent is running, waiting for an answer or finished. If that tracking does not save you time compared with your current tmux or terminal tabs, the rest of the feature list matters less.
4. Compare the diffs with the annotate feature: leave a comment on a line and send it back to the agent that wrote it. That feedback loop is the part that replaces copying text between windows.
5. Pair the phone only after the desktop flow works. The mobile app depends on the desktop host, and the Android release notes say most phone features come from the desktop, so keep both updated.
6. Check usage tracking. Run the same task across more agents only once you can see how much of your Claude or Codex allowance a single run consumes.
7. Review permissions for native chats, which now have a per-chat permission mode, before turning on Computer Use or letting agents use the Orca CLI to create workspaces.

If you already rely on skills, the shared-skills installer is worth testing next, because it installs one skill set across many agents instead of copying files into each agent's folder by hand.

What the Hacker News thread asked

The launch thread was small, with 21 points and 10 comments, so it is context rather than a verdict. The questions were practical: how it differs from existing tools such as cmux, how it reads status from GitHub, and whether the terminal layer could use a WebAssembly build of Ghostty instead of xterm, a suggestion from the commenter theturtletalks. The project's own answer to the terminal question is in the README, which describes a Ghostty-class terminal with WebGL rendering. Its answer to the comparison question is the scope: worktrees, mobile, tracking and tasks in one app.

Honest limitations and risks

1. Usage multiplies. Five agents on one prompt means five times the tokens and faster rate-limit burn. Orca's usage tracking helps you see it, but does not reduce it.
2. Experimental native chat. The release notes label it experimental and list known issues for remote hosts.
3. A wide surface. Computer Use, the Orca CLI and browser automation give agents more reach. Keep agents in worktrees, review permission modes, and avoid running unknown agent configurations against sensitive repos.
4. Telemetry. The README says Orca collects anonymous usage data and links privacy and opt-out docs; read them before installing on work machines.
5. Fast churn. Daily releases mean features and settings move; pin a version if you need stability.
6. Star counts are not quality. About 94.8k stars shows attention, not fit for your workflow.

If you let agents act on code or machines, an independent safety layer is worth having. AgentBeam, the agent security platform from the explainx.ai team, is built to stop AI agents before they take dangerous actions.

Should you try it?

Try Orca if you already juggle several agent sessions in tmux or multiple windows and want worktree isolation, notifications and a phone view. Skip it, or wait, if you use a single agent in one repo, or if you need a stable, slow-moving tool. Start with a throwaway repo, set explicit permission modes, and watch your usage for a day before scaling to five parallel agents.

Related reading

• cmux: terminal for AI coding agents
• AgentPlane: local control plane for Claude Code, Codex and Cursor
• OpenCode: the open-source AI coding agent
• Claude Code commands: complete reference
• ECC: 71 agents, 311 skills and what to check before you install

Sources: stablyai/orca README, Orca releases, and the Hacker News Show HN thread.

Version details and star counts are accurate as of October 17, 2026.