If you’ve settled into a single coding agent and gotten comfortable, Orca makes the case for running several at once. It’s an Agent Development Environment (ADE) built by Stably that treats “harness” as a plug-in concept rather than a single choice – you point it at Claude Code, Codex, Cursor, GitHub Copilot, OpenCode, Amp, and a long tail of other CLI agents, and it gives each one its own isolated git worktree to work in.
What it actually does
Orca’s core idea is simple: parallel agents need isolation to be useful. Instead of running five agent sessions in five terminal tabs and losing track of which one touched what, Orca spins up a worktree per agent, gives you a VS Code-style editor with autosave, and lets you drag files or screenshots straight into a prompt. You can comment on individual diff lines and ship that feedback straight back to the agent – review, edit, and commit without leaving the app.
It bills itself as bring-your-own-subscription: Orca doesn’t sell you tokens, it orchestrates whatever agent CLI you’re already paying for, tracking usage and rate-limit resets across providers so you can see at a glance which of your accounts is about to hit a wall.
Built for scale, not just convenience
A few features point at where Orca is heading. Its /orchestrate command delegates a task across multiple agents at once, and agents can drive Orca itself – scripting worktree creation, snapshots, clicks, and form fills as part of a workflow. There’s a headless Linux server mode (orca serve) for running fleets on a VPS, plus a mobile companion app so you can check on and steer agents from your phone. Rich previews for Markdown, images, PDFs, and repo docs mean you rarely need to tab out to see what an agent produced.
The project is genuinely popular – tens of thousands of GitHub stars and an active community pushing on its plugin architecture, including proposals to make agent harnesses themselves dynamically pluggable rather than hardcoded into the app.
Where it fits
Orca isn’t a harness in the sense that Claude Code or Cursor are – it doesn’t reason about your code itself. It’s the control layer above them: the thing you reach for once running one agent stops being the bottleneck and coordinating several becomes the actual problem. If you’re already comfortable with Claude Code or Codex individually, Orca is worth a look the moment you find yourself wanting to run two or three unrelated tasks against the same repo simultaneously without them stepping on each other.
Install via Homebrew on macOS (brew install --cask stablyai/orca/orca) or grab a build directly for Windows and Linux from the project’s GitHub page.
Leave a Reply
You must be logged in to post a comment.