Run AI coding agents across multiple computers
Chris Rupp
You can run agents on a desktop, laptop and build machine without an orchestrator. The practical problem is keeping track of which machine owns the files, which task is running and where its result landed.
Manual approaches
SSH and a persistent terminal: connect to a machine you administer, start the agent in a dedicated repository clone or worktree, and keep a session log. A terminal multiplexer can keep the session alive after your connection closes. This works well when you have a few machines and can manage assignment yourself. Restrict remote access; avoid exposing a shell directly to the public internet. Git documents working with separate worktrees.
A private network: a private networking tool can make machines reachable without publishing their services. It supplies connectivity; you still need to assign tasks, manage credentials and collect results. See Tailscale’s SSH documentation for one implementation.
Self-hosted CI runners: use a workflow to dispatch repeatable tasks to a machine with the required toolchain. Keep the agent’s credentials and repository permissions narrow. A runner is infrastructure you maintain, so review GitHub’s self-hosted runner security guidance before letting untrusted contributions execute there.
Containers or VMs: use these to bound the execution environment on each machine. They do not assign tasks across machines by themselves. Check mounted folders, network access and injected secrets; a container with broad host access defeats much of the purpose. Docker describes rootless operation as one way to reduce daemon and runtime privileges.
For any approach, give each task an explicit branch, acceptance checks and a publication destination. Push finished work before moving the same project to another computer. Avoid concurrent writes to a shared working folder.
Orbit multi-node: one queue across machines
Orbit’s Multi-Node Relay connects machines signed into the same Sivants account. You can queue work from one interface and choose the machine that executes it, with output streamed back. For example, keep a web project on a Windows PC and direct work requiring a Mac toolchain to your Mac.
Follow Set Up Multi-Node:
- Open Settings → Account → Connected Machines on each machine and sign into the same account.
- When queuing a prompt, choose its destination with Run on: [machine ▾].
- Set a project’s default execution target for routine work.
- Optionally select a primary machine for default node-local MCP tool routing.
Project authority still gives each project an owning machine, and prompt numbering is shared across the account. Each run uses its own isolated worktree. Keeping a project on one machine is the simplest arrangement; if you work on it from several, keep Git clean and pushed between sessions. Multi-node does not remove that coordination requirement.
Choose the smallest setup that fits
A remote shell is enough when you need one agent on another computer. CI runners fit jobs that already belong in a workflow. Orbit fits when you want to assign and follow queued work across projects and machines from one interface.
Before leaving a batch running, use the overnight safety guide. For queue control away from your desk, see the phone guide.