dbt MCP server

Lets an agent read your dbt project — models, lineage, metrics, job runs — query the Semantic Layer, and optionally run dbt commands like build, run and test.

Install

Run (pinned)
```
uvx dbt-mcp@2.4.0
```
Configure with environment variables — DBT_PROJECT_DIR and DBT_PATH for local dbt; DBT_HOST (plus DBT_TOKEN and DBT_PROD_ENV_ID for API access) for dbt Platform.

With an env file

uvx --env-file /absolute/path/to/your-dbt-project/.env dbt-mcp@2.4.0

Remote

dbt Platform also hosts a remote MCP server; see docs.getdbt.com/docs/dbt-ai/setup-remote-mcp.

dbt Labs’ own MCP server. It reads your dbt project’s metadata (models, sources, exposures, column-level lineage, model health and run history), queries metrics through the dbt Semantic Layer, generates staging models and YAML, searches the dbt docs, and can run dbt CLI commands and dbt Platform jobs.

When to use it

Answering “where does this number come from?” with lineage instead of guesswork.
Querying governed metrics so the agent uses the same definitions as your dashboards.
Writing or fixing dbt models with the project’s real schema in context.

Before you install

Start read-only. Set DISABLE_DBT_CLI=true and DISABLE_ADMIN_API=true until you trust the agent with builds and job runs, and point it at a dev target, not production. execute_sql needs a personal access token; service tokens don’t work for it.

Official

Published by the platform or the maker of the thing itself.

Access: Changes things freely Can create, change or delete in your accounts without asking each time.
Review: Reviewed — install with care
Scan: 100/100 SkillSpector 2.12.0 said do not install at 9efa046
Checked: Sep 25, 2026

2 CRITICAL and 69 HIGH, all read by rule and file. Every vulnerable-dependency finding (anyio, PyJWT, aiohttp, starlette, etc.) is in examples/aws_strands_agent/requirements.txt, not the server package. Credential hits are pydantic-settings reading .env and the OAuth token refresh flow for dbt Platform, which caches its context to a local file. The one curl-install is a CI step. "Hidden instructions" and "anti-refusal" are a PR template comment and the words "without warning" in docs. Two real behaviors earn CAUTION. First, the dbt CLI tools (build, run, clone, show) and Admin API tools (trigger, cancel, retry job runs) change warehouse objects and jobs; dbt's own README says so. Second, anonymous usage tracking is on by default and sends each tool call's name and arguments to dbt Labs, with sql_query and vars redacted. Set DO_NOT_TRACK=1 or DBT_SEND_ANONYMOUS_USAGE_STATS=false to turn it off. The scan now matches the pinned release. src/ at the v2.4.0 tag (9efa046) scores 100 with 0 CRITICAL and 21 HIGH, all "access token" and .env matches in the config and OAuth code plus two log strings; the whole repo at that tag gives the same 2 CRITICAL and 69 HIGH as before.

Publisher: dbt Labs
Source: github.com/dbt-labs/dbt-mcp
Works with: Claude apps, Claude Code, Cursor, Any agent
Categories: Data & analytics, MCP & integrations, Databases & warehouses
License: Apache-2.0
Cost: Free tier, paid plans

Curated by Andrew Borden-Chisholm. Notes on building with these tools are on the blog.