Run AI agent workers on any machine — git is your infrastructure
Multi-machine AI agent control plane using git as a distributed task queue. Atomic claim-via-push-rebase, Docker-isolated agent personas, Discord bot bridge, and Tailscale Funnel for remote trigger.
Distributed AI agent systems need coordination infrastructure — message queues, central servers, cloud dependencies. Hive-Claw eliminates all of that by using a git repository as the task queue: atomic task claiming via push-rebase, Docker-isolated personas per agent, and Tailscale Funnel for remote trigger — so any machine with Docker and git can join the swarm.
The challenge
Achieving race-condition-free task assignment across distributed workers with no central server — using only git's atomic push semantics.
The outcome
AI agent workers run on a home lab, a cloud VM, and a CI runner simultaneously — coordinated purely through git, with no message broker or orchestration service.
Technical approach
- Git-as-task-queue — branch-per-task model with atomic claim via push-rebase conflict resolution
- Atomic task claiming — git push-rebase as a distributed lock; only one agent wins the rebase
- Docker-isolated personas — each agent runs in its own container with independent filesystem context
- Tailscale Funnel — public HTTPS endpoint tunnelled to a local machine without port forwarding
- Discord bot bridge — webhook-driven trigger and status reporting via Discord API
- Actor model coordination — agents as independent workers that pull tasks, never share state
- Distributed locking via git — no central lock server; git commit graph is the authoritative state
- TypeScript Node.js daemon — persistent worker process polling the task branch for new work
Results at a glance
- Zero
- Infrastructure required
- Open source
- Source
