Deploy from Git with ChatGPT dots on xCloud

A dot does not wait in a chat tab for your next message. You write the goal once, in the ChatGPT desktop app or on desktop web, for example deploy github.com/example/shop to my Frankfurt server on a staging hostname and show me the dry run first, and the dot works from its own cloud computer. It calls git_detect through the xcloud plugin, reads the app type, branch and commands that xCloud reports, asks for a free staging hostname and sends the create request as a dry run. Then it comes back with the would_create summary instead of creating anything. Because a dot shares memory with ChatGPT, it can also remember that staging sites go on one server and production on another, and start from that the next time.

The approval is where a dot differs most from a chat. Two rules sit between the dry run and a live site. Yours come first: under Customize, Custom rules, a rule such as Ask before taking action for anything that creates or deploys on xCloud keeps the decision in your conversation, and Take action if pre-approved means you asked for that action in your prompt, so a sentence like deploy it if the dry run has no warnings is a real instruction. xCloud's rule comes second: servers_sites_git_auto is refused without an explicit confirmation. A dot that has your yes sends the confirmed request with an idempotency key, follows sites_status until the deployment is terminal and fetches the URL before it calls the site live. Choose Full access on the xCloud sign-in screen, because Read-only cannot create a site.

The dot connects from OpenAI's cloud, never from your computer, so it can only deploy what a Git provider holds. Unpushed commits and local folders are invisible to it. A private repository works when its provider is connected to your xCloud team or when you add the SSH deploy key that xCloud prepares. When a deploy fails, the dot reads sites_deploy-diagnosis, explains the cause and can retry on the same site with a corrected build command, but a change inside the repository still needs you or a coding agent to push it. A good standing instruction is a scheduled check: every weekday at 8, run detection and a dry run for one repository and message me any new warning. Detection is a read and the dry run creates nothing, but the dry run is still a write-scoped call: it needs a Full access connection, and whether it runs unattended is decided by your Custom rules rather than by xCloud. The create step waits for your reply either way.

ChatGPT dots specific: Check your plan before you promise yourself a hands-off deploy. OpenAI's help center says full MCP support with write actions is in beta on Business, Enterprise and Edu workspaces, and Pro accounts connect MCP apps with read and fetch permissions only. On Pro your dot can run detection and report on your servers, but do not expect it to send the create request. Dot access also depends on plan and region and can take days to arrive. A dot's proactive research cannot change content through plugins, so a deploy only ever starts from a task you gave it.