This isn’t a launch recap. Official GitHub docs for GitHub Copilot on GitHub.com and GitHub Mobile describe a Copilot in the cloud path where you can “ask questions and delegate work from a conversation,” and Copilot can “plan and make changes” and open work for you to review. The how-to for using Copilot in the cloud on GitHub says you can start a session from several places on GitHub, then “monitor its progress, steer it with follow-up prompts, and iterate on the resulting pull request without leaving the browser.” Here’s what the current official docs say.

Who the docs say can use it
The overview says that, starting October 28, 2026, “the new Copilot in the cloud experience on GitHub.com and GitHub Mobile is available to users with a paid Copilot plan, subject to applicable organization and enterprise policies.” “Users on Copilot Free plans will continue to use the old Copilot Chat experience.” The how-to says it’s “enabled by default for all paid Copilot plans,” and available “in all repositories stored on GitHub, except repositories owned by managed user accounts and where it has been explicitly disabled.” Repository owners “can disable Copilot for particular repositories,” and “organization and enterprise policies continue to apply.” “Without write access to a repository, your Copilot session is read-only.”
Why the docs say this is useful
The overview says Copilot can “research a repository, plan changes, and implement them in the background.” “It can edit files and run tests and linters in a cloud development environment, rather than your local IDE.” “When you ask for changes, Copilot works against the repository and branch you provided.” “You can review the proposed changes and ask for refinements before creating a pull request, or request a pull request in your prompt.” In both a cloud sandbox and a GitHub Actions-based session, “Copilot can only work on one branch at a time and can open one pull request per task.”
Start a session on GitHub.com
The how-to lists several entry points on GitHub itself: the Copilot home page at github.com/copilot, your signed-in personal dashboard at https://github.com, and the Agents tab or panel in a repository. In the prompt box, “describe what you want Copilot to do. Include the outcome you want and any constraints.” If you want a pull request, say so in the prompt. The official example is “Open a pull request to implement a user friendly message for common errors.”
Optionally, click Attach and choose files and folders, a repository, or a space if you want extra references. That’s different from choosing the working repository. Adding a repository through Attach “provides a reference (context material) for your prompt,” while selecting it under Work in a repository “makes the repository the working repository (scope) for the session.” “At the start of the session, the cloud sandbox takes a copy of the repository and gives Copilot permission, within existing constraints, to work with its content.” If a repository isn’t listed, “it may be excluded from Copilot by your personal settings or an organization or enterprise policy.”
Open the Mode picker to choose how the session collaborates. Interactive is the default: “Work with Copilot step by step.” Autopilot: “Delegate the task to Copilot for end-to-end execution.” In the same menu, Tool permissions control how tool requests are approved. Assisted is the default, and “uses a safety check to decide whether to approve a request or ask you.” Always ask will “require approval for each request.” Allow all will “automatically approve all tool, path, and URL requests.” A safety check is still a check, not a guarantee, and Allow all is the official option that skips those prompts.
Open Environment to choose where Copilot runs. Cloud is the default: “Run in a GitHub-hosted cloud sandbox.” Actions runs “in an GitHub Actions-based environment.” The cloud sandbox article says you choose the environment when the session starts, “and you can’t change it afterward.” Then click Send message or press Enter. The how-to says Copilot “will work on the task and push any code changes,” and “you can track Copilot’s work and open a pull request in one click from the session logs.”
What the docs say a cloud sandbox is
The sandbox article calls a cloud sandbox “an isolated Linux environment that GitHub creates for a Copilot session.” “When you work in one or more repositories, Copilot clones each one into the sandbox, then runs commands and tests and makes changes there.” “Because the session doesn’t run on your machine, you can run several sessions in parallel without using local resources, and you don’t need to stay connected while Copilot works.” “You don’t create or manage cloud sandboxes yourself.”
Each cloud sandbox “is isolated from your machine and from other sessions,” and “the commands that Copilot runs in a cloud sandbox can’t read or change the files on your computer.” That’s a useful boundary, not the whole risk picture. The official risks and mitigations article still warns that the agent “has access to code and other sensitive information, and could leak it, either accidentally or due to malicious user input.” GitHub’s documented mitigation is that it “restricts Copilot cloud agent’s access to the internet,” and network access “follows the same firewall settings as Copilot cloud agent, including the recommended allowlist and any custom allowlist rules that organization or repository owners add.” “Allowlist changes don’t apply to sessions that are already running.”
Cloud sandboxes “use your existing Copilot authentication.” The docs say “you don’t need to configure a separate cloud provider, manage API keys, or set up infrastructure.” That’s about not standing up a second cloud account for the sandbox, not a claim that there’s no authentication, and it doesn’t remove repository permissions. If Copilot comes through an organization or enterprise, “the Copilot in the cloud policy controls whether you can start sessions that run in a cloud sandbox,” and that policy “is enabled by default.”
Review the pull request, and don’t skip the workflow gate
The overview is blunt: “Copilot can produce incorrect or suboptimal code, including code that contains security vulnerabilities.” “Review and test the output before using it in production.” The risks article says the agent “can push code changes to your repository,” that “only users with write access to the repository can trigger Copilot cloud agent to work,” and that it “only has the ability to push to a single branch.” In the usual case, “a new copilot/ branch is created for Copilot.” “Draft pull requests created by Copilot cloud agent must be reviewed and merged by a human.” The agent “cannot mark its pull requests as “Ready for review” and cannot approve or merge a pull request,” and GitHub “prevents the user who asked Copilot cloud agent to create a pull request from approving it.”
GitHub Actions is a separate gate. The how-to and the cloud-agent settings page say that, by default, “GitHub Actions workflows will not run automatically when Copilot pushes changes to a pull request.” Those workflows “can be privileged and have access to sensitive secrets,” so “inspect the proposed changes” and “be especially alert to any proposed changes in the .github/workflows/ directory.” To let them run, click the Approve and run workflows button in the pull request’s merge box. An optional policy can drop that click, with this warning: “Allowing GitHub Actions workflows to run without approval may allow unreviewed code written by Copilot to gain write access to your repository or access your GitHub Actions secrets.”
By default, the agent “checks code it generates for security issues and gets a second opinion on its code with Copilot code review.” GitHub’s wording is that this “reduces the likelihood” of problems such as hardcoded secrets, insecure dependencies, and other vulnerabilities. The same page says those checks use CodeQL, the GitHub Advisory Database, and secret scanning. They’re GitHub’s built-in checks, not a substitute for review, and a repository administrator can disable them from repository Settings, under Copilot, then Cloud agent.
The official path is still a review path. A hosted sandbox keeps the agent’s commands off your machine, and the default workflow gate keeps Actions from running until someone with write access approves them. Neither one is a reason to merge unseen code. The hero photograph is Aqueduct at Blea Moor.jpg by The joy of all things, from Wikimedia Commons, licensed under CC0 1.0.