Orca hits 12,500 stars running dozens of CLI agents

Stablyai's Orca passed 12,500 GitHub stars running Claude Code, Codex, Cursor, Grok, OpenCode, Pi, and 30+ other CLI agents side by side. Each agent runs isolated in its own git worktree, steered from a desktop app with a mobile companion.

The billing model matters more than the fleet size. Orca runs on the subscriptions and API keys you already have. No per-seat markup stacked on top. For a solo builder, that turns buying an orchestration platform into running a tool on Tuesday.

Orca didn't rise alone. Cursor retired Composer for an Agents view that runs up to 8 parallel background tasks. Devin killed its Cascade mode for what it calls an Agent Command Center. Claude Code shipped background agents that commit, push, and open a draft PR without asking first, per Releasebot's report on the update. Four separate tools moved the same direction in one week. The unit of work shifted from one prompt editing one file to N agents fanning out across a repo while a human reviews what lands.

The catch nobody's advertising

Worktree isolation is what makes parallel agents survivable. It only works if your modules have real boundaries and your tests run fast enough to gate each agent's output on its own. Most repos don't have that. Modules that look independent turn out to share hidden state, so two agents editing separate-looking files still step on each other.

If you want to try this, don't install Orca first. Run git worktree by hand a few times on your messiest repo and watch where concurrent edits would collide. That collision map is your actual refactor backlog, and it'll tell you whether your codebase can handle a fleet before you hand it one.