dbt-mcp-server

mcp server that lets claude answer questions about your dbt project. model lineage, missing docs, test failures straight from manifest.json

stateless by design: every tool call re-reads target/manifest.json, so answers always reflect your latest dbt parse. core logic is stdlib only, with a single dependency on the official mcp SDK. talks over stdio.

requirements

• python 3.10+
• a dbt project with a generated target/manifest.json (any dbt command creates it)

installing

pip install dbt-mcp-server

configuration

hooking it up

claude desktop: add to claude_desktop_config.json (settings → developer → edit config):

{
  "mcpServers": {
    "dbt": {
      "command": "dbt-mcp-server",
      "env": {
        "DBT_MANIFEST_PATH": "/path/to/your/dbt_project/target/manifest.json"
      }
    }
  }
}

claude code:

claude mcp add dbt --env DBT_MANIFEST_PATH=/path/to/your/dbt_project/target/manifest.json -- dbt-mcp-server

tools

then just ask things like:

"what feeds into fct_orders, and what breaks downstream if i change it?"

"which models are missing docs? list the undocumented columns."

"summarize test failures since 2026-07-01."

note: the manifest has no true column-level lineage, so trace_column_lineage won't match renamed or derived columns; it flags upstream nodes without declared columns so you know where the trail could continue.

development

git clone https://github.com/vladrosant/dbt-mcp-server
cd dbt-mcp-server
pip install -e ".[dev]"
pytest

license

MIT