Connect your AI agent to Buddy with the new MCP server

You ask your AI assistant to deploy the latest version to production and to let you know if anything breaks. With Buddy's new remote MCP server at mcp.buddy.works it doesn't just reply, it actually runs the pipeline and tells you what happened.

What is MCP?

MCP is the protocol that allows an AI agent to actually do things, not just answer questions. A service exposes one MCP server and any compatible client, like Claude, Cursor or an IDE agent, can call it the same way, so you don't have to build a custom integration for every assistant. The server decides what's exposed and what's allowed.

Meet the Buddy MCP server

The Buddy MCP server is available at:

https://mcp.buddy.works/mcp

Use the endpoint matching your workspace region:

• US: https://mcp.buddy.works/mcp
• Europe: https://mcp.eu.buddy.works/mcp
• Asia: https://mcp.asia.buddy.works/mcp

It uses the modern streamable HTTP transport and is protected by OAuth 2.1 (authorization code + PKCE), so an agent only ever acts within the scopes you grant it. It exposes over 100 MCP tools covering the whole platform, the same actions you'd use in the Buddy UI, now available to an agent.

Connect it

Any client that supports remote MCP servers can connect - the setup config is below. There are two ways to authenticate:

• Interactive (OAuth) - the client opens your browser, you log in once and approve the scopes the agent may use. Best for desktop clients like Claude and Cursor.
• Headless (personal access token) - for CI, scripts, and custom agents, skip the browser and pass a Buddy personal access token as an Authorization: Bearer header instead (shown in the config below).

Create a token and pick its scopes in your Security settings. You can begin with read-only scopes while you're exploring, and add write access only once you actually want the agent making changes.

What your agent can do

The toolset maps directly onto Buddy's features:

• Projects & git - create and inspect projects, browse branches, commits, tags, and pull requests.
• CI/CD pipelines - read and edit pipeline YAML, start, cancel, retry, and approve runs.
• Sandboxes - spin up on-demand VMs, run shell commands inside them, move files, and snapshot their state.
• Packages & artifacts - manage the registry: create versions, upload and download files, publish whole archives.
• Deployments - inspect targets and configure distributions with routes that map domains and paths to per-region targets.
• Domains & DNS - list zones, check availability, and claim domains.

For example: you can ask it to spin up a sandbox from this branch and run the test suite. It creates the VM, runs the tests inside it, and tears everything down once they're done, streaming the output back to you along the way.

A note on safety

Some of these operations can't be undone: deleting a project, tearing down a sandbox, replacing a pipeline definition. Use OAuth scopes to control that risk. Start with a narrow, read-only token while you're still testing the workflow, then widen it once you trust the agent.

Try it

Add the server to your client. In any client that reads an mcp.json - Cursor, Claude Code, and others - point it at the endpoint:

json
{
  "mcpServers": {
    "buddy": {
      "url": "https://mcp.buddy.works/mcp"
    }
  }
}

In the Claude desktop and web apps, add it under Settings → Connectors → Add custom connector with the same URL. You complete the OAuth login once and approve the scopes.

For headless use (CI, scripts, custom agents), skip the login and pass a personal access token as a header instead (in a client that supports custom headers on remote servers):

json
{
  "mcpServers": {
    "buddy": {
      "url": "https://mcp.buddy.works/mcp",
      "headers": {
        "Authorization": "Bearer <YOUR_TOKEN>"
      }
    }
  }
}

Then give it a safe, read-only first task to confirm the connection, for example:

List my Buddy workspaces, then show the pipelines in <project> and the status of their last run.

From here you can widen the scopes and hand it real work: editing a pipeline, spinning up a sandbox, publishing an artifact. Your agent stops reporting on your delivery pipeline and starts running it.

• Buddy docs
• API & authentication
• What is MCP?