swarmagent.dev
GitHub
DocsModelsResourcesPricingNewsContactGitHub

Use casesAgent fleet

One agent fleet across all your machines

Install Swarm on every Linux machine you want to work on. Swarm's fleet MCP server finds them on your private Tailscale network and gives your AI one set of tools for all of them: the status of every machine, a job on the machine you name, and alerts from the whole fleet. Your Tailscale policy decides which devices may only read and which may give jobs.

Why more than one machine

  • Separate kinds of work: code on your PC, research and social posts on a VPS, builds on a build box.
  • Keep going when the laptop closes: servers keep running the queue.
  • Keep secrets where they belong: each machine has only the keys its own work needs.

How the fleet is wired

  1. Install Swarm on each machine (Ubuntu 24.04 x86-64). Your AI can do it over SSH.
  2. Put them on your Tailscale network with MagicDNS and HTTPS certificates on. Nothing is opened to the internet.
  3. Add the fleet MCP server to your AI once. It finds the machines by tag and reaches each one's tools.
  4. Name the machine in each call, or ask for the whole fleet's status at once.
Say this to your AI

Is everything all right across my machines? Then run the weekly dependency audit on the build box.

  • swarm_fleet_status
  • swarm_run_task machine=build-box

All 3 machines are ok; the VPS finished 4 jobs overnight for $0.31. The audit is running on the build box.

Watching the fleet

Each machine raises signals when a job fails, a worker is blocked or something needs you, and sends them to a webhook or Telegram. Your AI reads the same feed: what ran, what changed, what it cost. See Many machines and Alerts.

One model table for the fleet

Every machine runs its roles (coder, reviewer, finder, designer and the rest) on Swarm's recommended models for the provider you connected. You can override them for the whole fleet in one call, when you choose to.

Questions

Do I need Tailscale?

For the fleet MCP, yes: it finds and reaches the machines over your private Tailscale network. A single remote machine can also be reached over an SSH tunnel with no other account.

Can some devices only read?

Yes. Your Tailscale policy, or the key the device holds, decides whether it may only read the fleet or also give jobs.

Does each machine need its own model plan?

Each machine connects its own provider: a ChatGPT plan, a Claude plan through Claude Code, or an API key.

Is there a central server?

No. Each machine runs Swarm on its own; the fleet MCP is a client that talks to all of them.

Updated 24 October 2026