dbt-docs-mcp

Model Context Protocol (MCP) server for interacting with dbt project metadata, including dbt Docs artifacts (manifest.json, catalog.json). This server exposes dbt graph information and allows querying node details, model/column lineage, and related metadata.

Key Functionality

This server provides tools to:

• Search dbt Nodes:

  • Find nodes (models, sources, tests, etc.) by name (search_dbt_node_names).

  • Locate nodes based on column names (search_dbt_column_names).

  • Search within the compiled SQL code of nodes (search_dbt_sql_code).

• Inspect Nodes:

  • Retrieve detailed attributes for any given node unique ID (get_dbt_node_attributes).

• Explore Lineage:

  • Find direct upstream dependencies (predecessors) of a node (get_dbt_predecessors).

  • Find direct downstream dependents (successors) of a node (get_dbt_successors).

• Column-Level Lineage:

  • Trace all upstream sources for a specific column in a model (get_column_ancestors).

  • Trace all downstream dependents of a specific column in a model (get_column_descendants).

• Suggested extensions:

  • Tool that allows executing SQL queries.

  • Tool that retrieves table/view/column metadata directly from the database.

  • Tool to search knowledge-base.

Related MCP server: docmcp

Getting Started

1. Prerequisites: Ensure you have Python installed and uv

2. Clone the repo:

   git clone <repository-url>
   cd dbt-docs-mcp

3. Optional: parse dbt manifest for column-level lineage:

   • Setup the required Python environment, e.g.:

   uv sync

   • Use the provided script scripts/create_manifest_cl.py and simply provide the path to your dbt manifest, dbt catalog and the desired output paths for your schema and column lineage file:

   python scripts/create_manifest_cl.py --manifest-path PATH_TO_YOUR_MANIFEST_FILE --catalog-path PATH_TO_YOUR_CATALOG_FILE --schema-mapping-path DESIRED_OUTPUT_PATH_FOR_SCHEMA_MAPPING --manifest-cl-path DESIRED_OUTPUT_PATH_FOR_MANIFEST_CL

   • Depending on your dbt project size, creating column-lineage can take a while (hours)

4. Run the Server:

   • If your desired MCP client (Claude desktop, Cursor, etc.) supports mcp.json it would look as below:

   {
       "mcpServers": {
           "DBT Docs MCP": {
           "command": "uv",
           "args": [
               "run",
               "--with",
               "networkx,mcp[cli],rapidfuzz,dbt-core,python-decouple,sqlglot,tqdm",
               "mcp",
               "run",
               "/Users/mattijs/repos/dbt-docs-mcp/src/mcp_server.py"
           ],
           "env": {
               "MANIFEST_PATH": "/Users/mattijs/repos/dbt-docs-mcp/inputs/manifest.json",
               "SCHEMA_MAPPING_PATH": "/Users/mattijs/repos/dbt-docs-mcp/outputs/schema_mapping.json",
               "MANIFEST_CL_PATH": "/Users/mattijs/repos/dbt-docs-mcp/outputs/manifest_column_lineage.json"
           }
           }
       }
   }