External MCP servers

Give the Coding Agent tools and data from other services with remote MCP servers.

Connect an external MCP server when the agent needs tools from another service or your own system. To use pre.dev’s tools in another assistant, follow product MCP setup instead.

Add a server

1. Open MCP Servers in workspace Integrations, or the same tab in a project’s Project setup.
2. Choose Add server, then pick a preset or paste the server’s URL. pre.dev detects HTTP or SSE transport from the URL.
3. Name the server and add any headers it needs.
4. Test the connection and save.

Header values can reference your environment variables with ${NAME}, so a credential stays in one place:

{
  "url": "https://mcp.example.com/mcp",
  "headers": {
    "Authorization": "Bearer ${MY_API_KEY}"
  }
}

Advanced: edit config JSON accepts the same configuration directly.

Sign in to a server

Servers that use OAuth offer Connect. Sign in on the server’s own page; the card then shows who you are signed in as, with Reconnect and Disconnect. When a sign-in expires, the card asks you to reconnect. Each connect shows the host where you will sign in; if that host changes, continue only when you trust it.

Local commands run only in the CLI

Web builds use remote servers. A local command, such as npx -y <package>, only runs in the pre.dev CLI, on your machine. Add one there:

/mcp add files npx -y @modelcontextprotocol/server-filesystem ./docs

/mcp add <name> <url> adds a remote server from the terminal as well.

To let the agent use your own Chrome, with your open tabs and the sites you are signed in to, type /chrome. See Browser Agents Local.

Scope and controls

Personal projects inherit personal entries and can override a matching name. Organization entries are shared across the organization. See integration scope.

In the CLI, /mcp lists personal entries. Enter or Space toggles an entry, d removes it, and o opens workspace Integrations in the browser.

Enable the servers needed for the task, and describe the operation you want the agent to perform. A successful connection test confirms tool discovery; individual calls can still fail because of provider permissions or service errors.