CBGPU Deploy MCP

Ask your coding agent to set it up

Copy this into Codex or Claude Code on the computer where you use that agent:

Install and configure the official CBGPU deployment MCP for this coding agent.
Read https://github.com/CBGPU-Engineering/CBGPU-mcp/blob/main/README.md first.
Check this computer's OS, Node.js (22+) and available coding-agent CLI.
Use the pinned v1.0.2 GitHub release package from that README.
Register a USER-SCOPED stdio MCP named cbgpu-deploy for the coding agent I am using;
keep existing MCP settings. Use an absolute executable path if necessary.
Run the separate login command in an interactive terminal so I can enter my own
CBGPU email and password privately. Never ask for a password in chat, put one in
arguments/config, read an existing browser session, or print saved credentials.
If you cannot offer an interactive terminal, show me the login command to run.
Use the status command and MCP list_projects for frontend and server to verify.
Do not create, change, redeploy or delete projects during installation.
After setup, report whether login succeeded and whether the agent needs restarting.
When I later ask to deploy, inspect my repository and the existing CBGPU projects,
select the correct branch and root directory, then create or update that project.
Poll status and logs until it is ready or has failed; accepted/queued is not success.
Never copy local .env files or secrets into chat or MCP arguments.

CBGPU Deploy MCP lets Codex, Claude Code and other MCP clients deploy GitHub frontends, backend servers, Docker images and Docker Compose stacks through the same authenticated APIs as the CBGPU deployment console. The MCP runs locally over stdio; deployments run on CBGPU. The release bundles its runtime dependencies, so users do not need a separate MCP SDK installation. Bundled licenses are included in dist/THIRD_PARTY_NOTICES.txt.

Related MCP server: Nexlayer MCP

Connect manually

Requires Node.js 22+ and an approved CBGPU account with an email/password login. For a Google-only account, set an account password using CBGPU's password recovery flow first. This release does not implement browser OAuth login.

1. Sign in once

Run this in your own interactive terminal. The password is hidden and is never saved:

npx --yes --package=https://github.com/CBGPU-Engineering/CBGPU-mcp/releases/download/v1.0.2/cbgpu-deploy-mcp-1.0.2.tgz cbgpu-deploy-mcp login

2. Register with your coding agent

Codex:

codex mcp add cbgpu-deploy -- npx --yes --package=https://github.com/CBGPU-Engineering/CBGPU-mcp/releases/download/v1.0.2/cbgpu-deploy-mcp-1.0.2.tgz cbgpu-deploy-mcp

Claude Code:

claude mcp add --scope user cbgpu-deploy -- npx --yes --package=https://github.com/CBGPU-Engineering/CBGPU-mcp/releases/download/v1.0.2/cbgpu-deploy-mcp-1.0.2.tgz cbgpu-deploy-mcp

Restart the agent or reconnect its MCP servers after registering. On Windows, if the client cannot launch npx, use cmd /c npx as its command, or use the absolute path to npx.cmd. Windows users should restrict the session directory to their own account using the Windows file permissions dialog; Unix modes do not provide Windows ACL enforcement.

Other stdio MCP clients:

{
  "mcpServers": {
    "cbgpu-deploy": {
      "command": "npx",
      "args": [
        "--yes",
        "--package=https://github.com/CBGPU-Engineering/CBGPU-mcp/releases/download/v1.0.2/cbgpu-deploy-mcp-1.0.2.tgz",
        "cbgpu-deploy-mcp"
      ]
    }
  }
}

3. Ask for a deployment

Deploy this repository's frontend folder to CBGPU. Check existing projects first,
use the main branch and the frontend folder as root, and show me the production
URL after the deployment is ready. Do not expose environment secrets.

Deploy the server/ Docker Compose stack to CBGPU with 3 CPU cores, 4 GB RAM and
40 GB persistent storage. Use api-gw as the gateway service and container port
8000. Make it public at my-app.server.cbgpu.com. Keep volumes on redeploy and
inspect health diagnostics if a service fails.

Push local changes to GitHub before asking to redeploy: builds fetch the remote branch, not uncommitted files on your computer. Frontend production URLs use <slug>.front.cbgpu.com; server URLs use <slug>.server.cbgpu.com. Private GitHub frontends can use a repository token configured in the website. The MCP deliberately has no tool for reading or setting secrets.

Tools

Server settings.runtime is the legacy compatibility field. For newer runtimes, use settings.runtime_preset: node22, node20, node18, python312, python311, python39, java21, java17, go1, or dotnet10. settings.root_directory selects a subfolder for Git and Compose projects. Compose additionally accepts compose_file, gateway_service and storage_gb (40–200 GB). Existing website validation and capacity limits still apply. Some non-Compose projects do not expose runtime logs through the public API; this MCP reports that explicitly rather than fabricating logs or readiness.

Sessions and permissions

The website checks account approval, active login sessions and existing project owner/editor/viewer permissions for every request. No internal deployment key, service-role key or arbitrary shell access is given to the agent. A login is a normal CBGPU account session, not a separately scoped API token. It is stored locally outside repositories and automatically refreshed, with a cross-process lock so Codex and Claude Code can run together. The MCP only exposes the deployment tools listed above. Treat the session file as sensitive.

• Linux/macOS: $XDG_CONFIG_HOME/cbgpu-deploy-mcp/session.json, default ~/.config/cbgpu-deploy-mcp/session.json.

• Windows: %LOCALAPPDATA%\cbgpu-deploy-mcp\session.json.

• Unix directory/file permissions: 700 / 600; writes are atomic.

• Environment values and credentials are omitted from tool results. Logs remain untrusted user build output; do not print secrets in your application builds.

• Network timeouts do not automatically retry deployment mutations, avoiding accidental duplicate builds. Only a definitive authentication rejection is retried after refresh.

Check authentication or revoke only this MCP login, leaving browser/app sessions alone:

npx --yes --package=https://github.com/CBGPU-Engineering/CBGPU-mcp/releases/download/v1.0.2/cbgpu-deploy-mcp-1.0.2.tgz cbgpu-deploy-mcp status
npx --yes --package=https://github.com/CBGPU-Engineering/CBGPU-mcp/releases/download/v1.0.2/cbgpu-deploy-mcp-1.0.2.tgz cbgpu-deploy-mcp logout

CBGPU_ACCESS_TOKEN optionally supplies an existing short-lived user JWT instead of using saved login. It cannot refresh itself. Never put it in a committed MCP configuration. CBGPU_URL defaults to https://cbgpu.com and must be an HTTPS origin; changing it requires a separate login. CBGPU_ANON_KEY can override the public browser key for a separately operated CBGPU instance. This package contains only CBGPU's public publishable browser key, not a privileged server key.

Development

npm ci
npm run check
npm test
npm run build
npm run test:package
npm audit
npm pack --dry-run

Tests use the actual MCP SDK transport and cover tool routing, validation, Compose diagnostics, environment preservation, permissions, refresh races, private session storage, redirect refusal and logout. CI runs on Linux, macOS and Windows with Node.js 22 and 24. The packed release is installed into a fresh temporary directory and tested without an external SDK installation. No real user deployments are created by these tests.

Sources: MCP SDK, Codex MCP, Claude Code MCP.

License

Licensed under the Apache License 2.0.