Enhancement: add MCP (Model Context Protocol) server entry point

[stumelius]

Overview

The Model Context Protocol (MCP) is becoming the standard way for LLMs to call tools. Adding an MCP server entry point would let Claude Desktop, Cursor, and other MCP-compatible clients use dbt Cloud commands directly without any custom integration code.

Background

• dbt Labs has their own MCP server but it is currently unstable and not widely adopted
• dbt-cloud-cli already has all the building blocks: Pydantic command models, TOOL_REGISTRY, execute_tool_call (Add dbt_cloud/tools.py: auto-generate OpenAI/Anthropic tool definitions #135)
• An MCP server is essentially a JSON-RPC wrapper around get_anthropic_tools() + execute_tool_call()

Proposed approach

Add an optional mcp extra and a dbt-cloud-mcp entry point:

pip install dbt-cloud-cli[mcp]

The server would:

1. Expose tools/list → delegates to get_anthropic_tools()
2. Handle tools/call → delegates to execute_tool_call(tool_name, tool_input)
3. Read auth from env vars (DBT_CLOUD_API_TOKEN, DBT_CLOUD_HOST)
4. Respect DBT_CLOUD_READONLY to expose a read-only subset

Dependencies

• mcp Python SDK (optional dependency)

Related

• RFC: First-class support for agentic / autonomous agent use #107 — parent RFC
• Add dbt_cloud/tools.py: auto-generate OpenAI/Anthropic tool definitions #135 — tools.py (provides TOOL_REGISTRY and execute_tool_call)
• Add DBT_CLOUD_READONLY env var to restrict destructive commands #134 — DBT_CLOUD_READONLY