MCP servers

Expose your own tools, databases, and APIs to the chat over the Model Context Protocol.

The Model Context Protocol is a small, well-specified way for an LLM to call external tools. NBI manages MCP servers from inside the lab — add, remove, enable, disable — and routes their tools into the agent’s toolbox.

What you get

  • In-lab management. A dedicated panel in Settings lists every MCP server, transport (stdio, SSE, HTTP), and status. Add a new one without dropping to a terminal.
  • Per-server enable/disable. Toggle a server off without uninstalling it, useful when iterating on a flaky local server.
  • HTTPS-only for remote. SSE and HTTP transports require https:// URLs; flag-smuggling guards reject malformed env / header inputs at add-time.

Add a server

From Settings → MCP, click Add server and provide a name, transport, and command (for stdio) or URL (for SSE/HTTP). Example for the official Postgres MCP server:

{
  "name": "postgres",
  "transport": "stdio",
  "command": "npx",
  "args": ["-y", "@modelcontextprotocol/server-postgres", "postgresql://localhost/analytics"]
}

For an HTTP/SSE server hosted internally:

{
  "name": "internal-search",
  "transport": "http",
  "url": "https://mcp.internal/search",
  "headers": { "Authorization": "Bearer ${MY_TOKEN}" }
}

Reference

Install NBI