terradart-mcp

terradart-mcp is a Model Context Protocol (MCP) server that exposes TerraDart’s curated Google Cloud factory catalog to coding agents. The catalog covers 373 entries (372 curated resource factories + 1 data source) across 61 service barrels (compute, container, pubsub, cloud_run, alloydb, redis, iap, firestore, and more). With the server connected, an agent can look up the exact constructor signatures, nested types, and ready-made Stack templates it needs to author correct TerraDart Dart code — instead of guessing factory names from memory.

It is built with genkit_mcp (Genkit’s MCP server library) and ships as a single compiled binary. Coding agents connect via Claude Code, Claude Desktop, or Cursor; a Genkit Dart app can host the server as an MCP client too (Connecting clients).

How it fits together

graph LR
  agent["coding agent<br/>(Claude Code / Cursor / Claude Desktop)"]
  mcp["terradart-mcp<br/>list_barrels · list_resources<br/>get_resource_schema · get_quickstart<br/>check_coverage"]
  catalog["static catalog in terradart_google<br/>334 entries · 59 service barrels"]
  agent -->|stdio JSON-RPC MCP| mcp
  mcp -->|reads in-process| catalog

The agent speaks JSON-RPC over stdio. terradart-mcp answers each tool call by reading the catalog that is compiled into terradart_google — no separate data files, no network round-trips.

What it does NOT do

terradart-mcp is read-only catalog metadata. It only answers questions about TerraDart’s curated factories. It does not:

• run terraform (no plan, no apply);
• touch Google Cloud, your project, or any credentials;
• write, synthesize, or apply infrastructure;
• need network access to do its job.

If an agent wants to actually synthesize and apply a stack, that still happens the normal way — see How it works. terradart-mcp exists purely so the agent writes the right Dart in the first place.

Next steps

• Install — Homebrew or a direct binary download.
• Connecting clients — Claude Code, Claude Desktop, Cursor, and Genkit Dart.
• Tools reference — the 5 tools, with request/response examples.
• Recipes — prompt patterns for discovery, schema lookup, scaffolding, and plan review.