MCP Server (AI Editor Integration)¶

Docglow includes a Model Context Protocol server that exposes your dbt project to AI editors like Claude Code, Cursor, and Copilot.

Setup¶

Add to your editor's MCP config (e.g. ~/.claude.json for Claude Code):

{
  "mcpServers": {
    "docglow": {
      "command": "docglow",
      "args": ["mcp-server", "--project-dir", "/path/to/dbt/project"]
    }
  }
}

The server runs locally over stdio. No API keys or network access required.

Available Tools¶

The MCP server exposes 9 tools:

Tool	Description
`list_models`	List all models with metadata (name, description, materialization, folder)
`get_model`	Get full details for a model (columns, tests, SQL, dependencies)
`get_source`	Get full details for a source (columns, freshness status)
`get_lineage`	Get upstream/downstream dependencies for a model
`get_health`	Get the project health report (scores, coverage, violations)
`find_undocumented`	Find models and columns missing descriptions
`find_untested`	Find models and columns without tests
`search`	Full-text search across models, sources, and columns
`get_column_info`	Search for a column name across all models in the project

Use Cases¶

With Claude Code or Cursor:

"What models depend on stg_orders?" — the AI uses get_lineage to trace dependencies
"Which models need documentation?" — uses find_undocumented to list gaps
"What does the customer_id column mean across the project?" — uses get_column_info
"Help me write a description for dim_employee" — uses get_model to understand the model's SQL and columns

In CI/CD:

# Start the MCP server for automated documentation tasks
docglow mcp-server --project-dir /path/to/dbt

Options¶

docglow mcp-server --project-dir /path/to/dbt    # Required: dbt project root
docglow mcp-server --target-dir /path/to/target   # Optional: custom target directory