What actually happened
A Claude Code cloud session runs on Anthropic-managed infrastructure, or on a self-hosted environment if the organisation has one. It keeps working after you close the laptop. You can start one from claude.ai/code, the Code tab in the Claude mobile app, Cloud in the desktop app, or the terminal with claude –cloud.
Each –cloud command is its own session and its own branch. Anthropic’s own example is three jobs at once: fix a flaky auth test, update API docs, refactor the logger to structured output. When a session finishes you can open a pull request from the browser or bring the branch and conversation into your local terminal with claude –teleport. Teleport insists on a clean working tree, the same repository, and the same claude.ai account.
Cloud sessions are on Pro, Max and Team, plus Enterprise premium or Chat + Claude Code seats. They share your existing Claude usage limits. There is no separate charge for the VM. Clone and push need GitHub. GitLab or Bitbucket can be sent as a local bundle, but the session cannot push back to those remotes.
Why a small team cares
Take a three-person agency on a Thursday. One client WordPress plugin is throwing errors. Another client wants API docs brought up to date. A React app needs a dependency bump that does not touch the first two repos.
Those jobs do not share a branch. Previously one developer worked them in sequence, even with an assistant in the local terminal. The machine had to stay awake. The limit was one pair of hands.
Cloud sessions change the unit of work. You write a tight brief, point the session at the repo, and go to a client call. You check progress from the phone. You teleport when you want the diff on your own machine.
The useful question is no longer how fast Claude finishes one task. It is how many well-bounded tasks you can supervise without spending the afternoon untangling pull requests. Documentation, an isolated test fix and a logging change are good candidates. Three agents inside the same authentication module are not.
Hype vs useful
This is not a junior hire. Parallel agents do not make a tangled codebase parallel.
The hype version is ten sessions overnight and a clean main branch in the morning. The useful version is three independent jobs, each with a test Claude can run, each producing a pull request you can read in minutes.
Review is the constraint. Launch ten tasks and you replace a coding queue with a review queue. The metric that matters is accepted work per hour of human attention, not session count.
Access is the other constraint. A docs task does not need production database credentials. A frontend refactor does not need Stripe. Saved cloud environments control network access, environment variables and a setup script. Use the tightest environment that can still run the tests.
Rate limits are shared with the rest of your Claude usage. Three parallel sessions burn the allowance faster. They do not add a VM invoice, but they can exhaust the plan before Friday if you treat them as free compute.
If I were running this shop I would not make cloud sessions the default for every ticket. I would park independent backlog work there, keep coupled architecture work local, require each session to run tests, and measure review time before I added a fourth session.