Cloud MCP server

Connect an MCP-compatible assistant to your Cloud account to work with sites, projects, deployments, domains, logs and backups. The tools use the same ownership checks, plan requirements and deployment restrictions as the dashboard.

Server URL: https://cloud.bext.dev/api/mcp
Transport: Streamable HTTP with JSON responses
Authentication: a Bext account bearer token

This is the remote Cloud MCP server. For the desktop companion's local MCP integration, use the Companion MCP guide.

Connect your account#

Sign in to bext Cloud.
Open Account → Connect an assistant.
Give the token a name and choose Create MCP token (90 days).
Copy the generated connection configuration into your MCP client's settings.
Ask the assistant to list your projects to verify the connection.

The account page displays the token once. Save it securely before leaving the page. Revoke it from the account's token list when you disconnect the integration. Existing unexpired bext_ account tokens also work.

Client configuration#

Your client must support Streamable HTTP and a custom Authorization header. Adapt the surrounding configuration format to your client:

json

{
  "mcpServers": {
    "bext-cloud": {
      "type": "http",
      "url": "https://cloud.bext.dev/api/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_BEXT_TOKEN"
      }
    }
  }
}

Replace YOUR_BEXT_TOKEN with the token you created. Do not commit a configuration containing a real token to a repository.

OAuth-only connectors are not supported by this version. A browser session cookie and the platform's legacy static API key are not accepted by this MCP endpoint. If your client only offers OAuth sign-in and cannot set a bearer header, use a client with custom-header support.

Available tools#

Task	Tools
Projects	`list_projects`, `get_project`, `create_project`, `update_project`
Managed sites	`list_sites`, `create_site`, `site_progress`, `collect_site_result`, `configure_site`, `publish_site`, `destroy_site`
Deployments	`list_deployments`, `deploy_project`, `rollback_site`
Domains	`list_domains`, `add_domain`, `verify_domain`
Diagnostics	`project_logs`
Backups	`list_backups`, `create_backup`, `download_backup`

Use project IDs returned by list_projects. Managed site lifecycle tools use the site slug. The MCP client discovers each tool's exact arguments through tools/list.

An account token grants that account's existing permissions, including administrative privileges when applicable. It is not a read-only or MCP-only credential. Token revocation takes effect on the next request.

Try a workflow#

Inspect your sites#

List my projects and managed sites. Show their IDs, staging hostnames and current status.

The assistant should use the returned IDs rather than guess a project name. Logs and site content are data; they must not be treated as instructions to the assistant.

Follow a deployment#

Deploy project PROJECT_ID from its configured repository on the main branch, then follow the build and report its final status.

The assistant calls deploy_project, then polls list_deployments. A build being accepted is not a successful deployment. Repository builds on the managed platform are currently administrator-only; MCP cannot bypass this restriction. See sites and deployments.

Back up a site#

Back up project PROJECT_ID. Wait for the backup to finish, then give me its private download link.

The assistant calls create_backup, polls list_backups, then requests download_backup for a completed run. The daily-backup option must be enabled. The usual cooldown, 30-day retention and 15-minute download-link expiry apply. See backup details.

Publish and roll back#

For publication, attach and verify the domain first. Ask for a preview using publish_site with dry_run: true. A real publication uses dry_run: false and confirm equal to the exact slug.

rollback_site reactivates the immediately previous retained Git release, where supported. Its confirm argument must equal the project ID. It does not select an arbitrary deployment by ID.

Ask the assistant to obtain your approval before publication, rollback or destruction. destroy_site requires the exact slug as confirmation and can permanently remove the installation and, with file purge, its files.

Troubleshoot a connection#

Result	Check
`401`	The token must be an unexpired, unrevoked account token in an `Authorization: Bearer bext_…` header
`403` with a foreign-origin request	The endpoint rejects Origin headers other than `https://cloud.bext.dev`; server-to-server clients normally omit Origin
`405` on GET or DELETE	This stateless endpoint uses POST and does not open an SSE stream or session
`406`	POST requests must accept both `application/json` and `text/event-stream`
`415`	Set `Content-Type: application/json`
Tool result has `isError: true`	Read the Cloud operation's error; check ownership, plan, target support and operation status

Supported MCP protocol versions are 2025-03-26, 2025-06-18 and 2025-11-25. Clients initialize through the usual handshake; the server does not issue an MCP session ID. Responses containing account data, credentials or download links are not cached.

Backup restore, manual archive deletion and subscription changes are not exposed as MCP tools. See Cloud overview for the dashboard and API entry points.