Skip to content

Tools

Copilot gains cross-repository projects and sessions that follow you across devices

Edited and reviewed by AI Ping Editorial Team

3-line summary

  1. Projects and desktop computer use arrive in public preview.
  2. A new paid-plan cloud experience connects VS Code, web, mobile and the app.
  3. Cloud runtime billing begins in November and is managed separately from AI credits.
PracticalDevelopersOffice workers
Code printouts and specification pages beside a laptop and phone
AI-generated conceptual image: code, specifications and devices gathered in one project workspace · AI Ping

GitHub introduced Projects in the Copilot app in public preview on October 28, bringing work across repositories into a shared goal with its supporting material and results. It also began rolling out a new cloud experience to paid Copilot users. The same cloud session can continue across the app, VS Code, the web and mobile.

A project can include several repositories or start with documents alone. Users describe a goal and constraints, and Copilot coordinates related work. This puts repository changes alongside planning material that sits outside code. Users can still start a focused session directly from a particular repository. GitHub also described using connected Slack and Figma context to investigate requests and draft specifications, alongside image generation and editing in chat.

Progress is grouped into items needing input, open pull requests, working sessions and completed work. Users can inspect a session’s conversation, changes and plan, then send further direction. The project Library distinguishes material supplied by the user from artifacts generated by agents.

Automations can run locally or in the cloud. GitHub’s app documentation says cloud automation prompts and session logs are visible to anyone with read access to the repository, including personal automations. A per-run AI-credit limit stops an execution when reached. Daily, weekly and monthly limits prevent new runs, but allow an existing run to continue.

A continuous conversation, with separate compute costs

The new cloud experience requires a paid Copilot plan; Free users retain the earlier web and mobile chat. Cloud sessions run commands and tests in isolated Linux environments managed by GitHub. Work can continue without a connection from the local computer. Stopped sessions retain snapshots of files and work in progress for later resumption.

The execution environment is chosen when a session starts. Changing it requires a new session. Repositories relying on Actions-specific setup steps, self-hosted runners or a Windows development environment need the Actions option. GitHub’s comparison describes Actions-based changes as confined to one repository, while cloud sandboxes can work across repositories within an organization.

Organization policies and repository exclusions carry over. Access to organization session data requires its single sign-on requirements; sessions without repository write access are read-only. The comparison lists 28-day chat retention for Actions-based sessions and account-lifetime retention for sandbox sessions. Session logs record the work and tools used.

An illuminated server behind a closed laptop and phone
AI-generated conceptual image: a remote server represents cloud work continuing beyond a closed laptop · AI Ping

Cloud sandbox usage billing begins in November 2026. Included Actions compute becomes a shared monetary allowance. Running time, allocated memory and snapshot storage for stopped sessions are measured separately. Compute and memory accrue while running; memory is measured by its allocation rather than active consumption.

Sandbox budgets are separate from AI-credit budgets. GitHub’s billing guidance supports product-level or individual usage-item budgets. Enabling the stop-usage option blocks further use when a budget reaches its limit. For sessions working in repositories, costs are assigned to the account that owns those repositories.

The app also gained computer use in public preview on macOS and Windows. In local sessions, it can read application content, click controls, enter text and move between apps, including desktop software without a dedicated integration. It is disabled by default, and enterprise administrators can block it through managed settings.

Approval follows the session’s tool-permission settings. An Always allow decision is shared by the app and CLI on the same computer. Removing it affects future approvals; access already granted to an active session must be stopped separately.

A hand over a pause key in front of a laptop
AI-generated conceptual image: a user pausing an agent working with desktop applications · AI Ping

Macs also require Accessibility and Screen Recording permissions. GitHub documents limitations including wrong-control selection and repeated actions when window state or timing changes.

Sources