What is a cloud agent?
A cloud agent is a coding agent that runs on a hosted machine instead of your laptop. You hand it a task from a browser, phone or terminal; it clones your repository into an isolated VM, works while you do something else, and pushes a branch or opens a pull request for you to review.
Part of Glossary: the terms behind MCP, coding agents and shared pages, defined plainly.
You start a task from the web, a desktop app, a phone, the terminal (claude --cloud) or a Slack or GitHub mention, depending on the product.
The service provisions an isolated VM, clones the repository from GitHub at your branch, and runs your environment's setup script or restores a snapshot.
The agent reads, edits, builds and tests inside that VM. You can close your laptop; the session keeps running.
The environment decides what the VM can reach: an allowlist of domains, secrets injected at start, git credentials held outside the sandbox by a proxy.
When it finishes it pushes a branch. You read the diff, comment on lines, ask for changes, open a pull request, or pull the session into your terminal.
Cloud agent versus local agent
A local coding agent runs in your terminal or editor, on your files, with your credentials and your network. It is fast to steer and sees exactly what you see, including the half-finished edit you have not committed. A cloud agent runs on a machine the vendor provisions. It sees only what it can clone, reaches only the domains its environment allows, and reports back through a web page, a phone app or a pull request.
That trade is the whole point. Because the agent is not on your laptop, you can run many at once, walk away, and come back to branches. Because it is isolated, a mistake it makes stays in a disposable VM. The costs are a slower feedback loop and a setup step: the VM needs your dependencies installed before the agent can build or test anything, which is why every product has a notion of an environment.
How the three main products do it
Claude Code calls them cloud sessions. Each one runs in an isolated, Anthropic-managed VM. You start from claude.ai/code, the Claude app, the desktop app or claude --cloud "fix the flaky test" in a terminal, and each command creates its own session, so three commands run three tasks in parallel. GitHub credentials stay encrypted on Anthropic's servers; the VM talks to GitHub through a proxy that attaches them. When a session finishes you can open a PR from the web or run claude --teleport to fetch its branch and conversation into your terminal. It counts against your plan's usage, with no separate charge for the VM.
Cursor's Cloud Agents, formerly Background Agents, also run in isolated VMs with full development environments. They clone from GitHub, GitLab, Azure DevOps or Bitbucket, and you can start one from the desktop app, cursor.com/agents, the iOS app, Slack, Linear, or by commenting @cursor on a PR. Environments come from a Dockerfile in .cursor/environment.json, a snapshot, or agent-led setup, and each agent pushes to its own branch, often with screenshots or a video of what it changed.
Codex cloud gives each task its own workspace that keeps working while your computer is asleep. You connect GitHub, let Codex inspect the repositories and publish an environment, then start tasks from ChatGPT on the web, mobile or desktop, or from the Codex CLI. Internet access is a per-environment preset with additional allowed domains and network secrets.
When to use it
Send to the cloud the tasks that are well specified and slow: a migration you already planned, a batch of lint fixes, a flaky test, documentation that follows the code. Plan locally first if the task is fuzzy; Claude Code's docs suggest plan mode on your machine, committing the plan, then executing it in the cloud. Keep on your laptop anything that needs your uncommitted work, your local services, or a tight back-and-forth.
Common mistakes
- Starting before pushing. The VM clones the remote at your branch, not your checkout. Commits only on your machine are not there.
- Adding a secret after launch. Cursor injects secrets at startup, so a running agent will not see a new one. Start another.
- Ignoring the network setting. Hosted environments limit outbound access by default, and a build that needs a private registry fails until you allow it.
- Forgetting shared rate limits. Parallel sessions draw on the same account limits as everything else you run.
FAQ
- Is a cloud agent the same as a background agent?
- Mostly. Cursor used to call them Background Agents and now calls them Cloud Agents; Codex runs them as tasks in the cloud from ChatGPT; Claude Code calls them cloud sessions. All of them run the agent on a hosted machine so it keeps going while you do other work. A local agent you leave running in a terminal is also in the background, but it is still on your laptop.
- Can a cloud agent see my uncommitted changes?
- Usually not. The VM clones your repository from GitHub at your current branch, not your working copy, so push first. Claude Code can upload a local bundle when there is no remote, and it leaves uncommitted edits to files that look like credentials or keys behind on macOS and Linux.
- How does a cloud agent get access to my repository?
- Through a source-control connection you grant once, typically a GitHub App installed on the repositories or your own token sent to the service. In Claude Code's hosted environments the credential never enters the VM; a proxy adds it to each GitHub request on the server side.
Stop sending HTMLover Slack.
Create an account in a minute, connect your coding agent, and make your first board.