Get started

Introduction

Coding agents have got good. Good enough that the hard part isn't the agent anymore, it's everything around it. Where does it run? How do you keep five of them going without your laptop turning into a space heater? How do you check on one from the queue at the coffee shop?

Sheepdog is our answer. You give an agent a task. It goes off and works on a cloud machine of its own. You check in when you feel like it, from your laptop, the desktop app or your phone.

How it works

You start a task by writing a message, the same way you'd ask a teammate for something. Sheepdog spins up a fresh machine, clones your repos, runs your setup script, and hands the work to Claude Code or Codex.

A task is just a conversation. You can reply whenever you like, and the agent reads your message next. You don't have to wait for it to finish before changing your mind.

When it's done, you get a notification. Then you do whatever you'd do with a teammate's work: read the pull request, try the change in a preview, or ask for something else.

Workspaces

Most of the work of using an agent is setup. It needs the right repos, the right environment variables, the right tools. Doing that once per task gets old fast.

So you do it once per project instead. A workspace holds a project's repos, setup script, env vars, integrations (MCP servers) and skills, and every task you start in it gets all of them.

You can keep a workspace to the people who need it, or open it up to your whole organisation.

Tasks and machines

Every task gets its own machine. They never share, which means you can run as many as you want side by side and none of them will trip over each other's files or ports.

Machines pause after 10 idle minutes, so they aren't sitting there running when nobody needs them. Send a message and it wakes back up with its files exactly as they were.

Agents can also show you things. When one starts a dev server, you get a preview link you can open in a click, though only people who can see the task can open it. They can drop screenshots, recordings and reports straight into the conversation too.

Your subscriptions

You bring your own Claude Code and Codex subscriptions, or API keys. If you're already paying for one, you don't need to pay for another.

If you add more than one, Sheepdog moves on to the next when one hits its usage limit. Your tasks keep going instead of stalling until the limit resets.

Working as a team

Agents are more useful when other people can see what they're doing. Share a task with a teammate, or share the whole workspace, and they can follow along and send it messages of their own. Everyone who's written in a task hears when the agent finishes.

Next steps

The best way to get it is to try it. Sign in and give an agent something small to start with.

If you're on a Mac or Linux, get the desktop app. And if you want agents to reach things that only live on your computer, like a simulator or a local database, read about Devices.