GitHub Copilot Cloud now spans devices – but shares the Actions compute budget

GitHub now connects sessions across github.com, GitHub Mobile, Copilot CLI, the Copilot app, VS Code and some Slack or Teams flows. People can start a coding task in one place and resume it elsewhere. This is more than sharing a chat transcript, but it does not demonstrate identical behavior in every local development environment.

The vendor says connected cloud sessions and sandboxes are generally available on paid Copilot plans. Actual repository access and tool availability remain governed by settings, membership and policies.

Cloud sessions run by default in GitHub-managed sandboxes, with snapshots intended to support resumable tasks. These properties may help longer-running work, but they do not independently prove that secrets, network connections or every external state can be transferred without risk.

Production teams should inspect repository and credential access, tool approval, pull request review and the behavior of resumed sessions. AgentenCode did not execute an independent sandbox security evaluation.

GitHub says the included Actions entitlement is now a shared dollar-based compute allowance. Standard hosted runners and cloud sandboxes use it; eligible larger hosted runners also participate on Team and Enterprise Cloud. This is a compute entitlement, not a universal Copilot token budget.

Cloud sandbox billing begins in November 2026. Each compute product retains its own published rate. Without actual runtime, plan and usage data, any precise price per agent job would be an invention.

GitHub describes a “Copilot in the cloud” organization policy, repository exclusions, enterprise MCP allowlists and tool approval requirements. Moving a session from a laptop to a phone does not confer unrestricted tool or repository access.

Our editorial recommendation is to start in a test repository, restrict identities and secrets, explicitly allow only necessary connectors and require human review of code changes. A vendor-managed sandbox is not an independent security certificate.

The seven-row source-backed matrix separates availability, clients, sandboxes, compute, billing, policies and missing independent tests. It is a practical management checklist rather than a copied changelog.

Our original illustration follows a task through device choice, the sandbox, Actions compute and approval checkpoints. It is an analytical diagram, not a real billing dashboard screenshot.

A useful first test is a small maintenance task in a noncritical repository. Record permission prompts, the proposed pull request, CI runs, sandbox usage and the relevant budget display. Make the code review explicit.

Then test resumption across devices. Confirm context, repository permissions and task state remain appropriate. Without this run, any assertion about cross-device reliability remains unverified.

Connected Copilot cloud sessions may make coding agents more convenient. The distinctive governance change is that agent sandbox compute and CI consume a shared included value. Organizations need a clear budget owner and review process.

Our recommended sequence is to restrict access, run a measured pilot, inspect costs and proposed code, and only then expand concurrency. Our earlier Copilot Automations article covers shared triggers and approval concerns.