Project Swarm

Verrow’s open-source build system. One coordinator splits the work, a swarm of agents does the scoped pieces, and nothing merges until every check is green.

GitHub ↗

How a run works

  1. Plan. The coordinator turns the goal into a task file and splits it into scoped jobs. Any agent can coordinate: Claude Code, Codex CLI, Cursor or Gemini CLI.
  2. Route. Each job goes to the cheapest lane that can do it. Jobs set to auto move up a lane after two failed or timed-out attempts.
  3. Work. Workers only see the files they were given. Writers run in their own git worktree under a sandbox.
  4. Review. The coordinator reads every result and integrates it into the project.
  5. Merge. The merge guard refuses anything unless every check is green.
Watch them leave, work, come back for review and merge.
No jobs dispatched yet.

Worker types

Hover a row to pick those workers out of the swarm.

TypeAgentsWhere it runs
Worktree writercodex, cursorOwn git worktree, macOS sandbox
Claude shellclaude + shellOwn git worktree, macOS sandbox
Copied workspaceclaudeA copy of the files it was given
Tool-free CLIhermes, qwenCopied workspace, JSON exchange
Tool-free APIopenai, gemini, ollama, lambda, openrouterOne HTTP request

Robot heads: robot-heads by Fayaz Ahmed, MIT licence.

Lanes

Cheapclaude · haikuSmall, mechanical jobs
Midcursor · composer-2.5Falls back to claude · sonnet
ExpensivecodexFalls back to claude · opus

Lanes are routing settings on each machine. A job that names its own agent and model keeps them.

Built to be trusted

  • Merges go through swarm merge, which refuses unless every check is green.
  • Writers work in a per-job git worktree under a macOS sandbox, so a bad job can’t touch the rest of the repo.
  • Kickoff asks which model providers may see the code and what the spend ceiling is before any model call.
  • Workers never see secrets.
  • Every problem from a real run is written down as a lesson and, where possible, pinned with a check and a test.

By the numbers

v1.167.37latest version
214versions
November 12, 2026last commit

Live from GitHub, refreshed hourly. Apache 2.0.

Start a project with it

Paste this into the agent you want to coordinate. The full kickoff prompt is in the repo.

Use Project Swarm for this project. Read docs/kickoff.md in the toolkit and perform its kickoff workflow. Ask me up front which model providers may receive project code and what spend ceiling applies; wait before model calls.
Open on GitHub ↗Read the case study

Want it set up on your project? Email braden@verrow.ai or call (870) 365-6644. More on how we build agent systems: AI agent services.