@us-all/dbt-mcp

dbt MCP server — manifest.json, run_results.json, sources.json, catalog.json, plus DQ result tables (BigQuery / Postgres) behind one stdio MCP. Built on @us-all/mcp-toolkit.

A read-only window into your dbt project for LLM clients. No dbt run triggering — just deep introspection, run-history analysis, source freshness, per-column test coverage, lineage walks, and (if you have a custom DQ result table) historical check trends and Tier SLA status.

For DAG triggering / run history / log tails, install the companion @us-all/airflow-mcp alongside.

• 27 tools across 3 categories (dbt, quality, meta) — 21 primitive tools + 5 aggregations + 1 meta
• 4 MCP Prompts for triage workflows
• 5 aggregation tools that replace 3-5 round-trips of "list / get / list"
• extractFields response projection on high-volume reads
• Read-only by default
• Hybrid backend: BigQuery (default) or Postgres for DQ result tables — both peer-imported lazily

Install

# 1. add the MCP server
pnpm add -D @us-all/dbt-mcp
# 2. add the DQ backend you actually use (only if you query custom DQ tables):
pnpm add -D @google-cloud/bigquery   # OR
pnpm add -D pg

Run

DBT_PROJECT_DIR=/path/to/dbt-project \
DQ_RESULTS_TABLE=my-project.data_ops.quality_checks \
npx @us-all/dbt-mcp

The server speaks MCP stdio; wire it into Claude Desktop / Cursor / any MCP client. Set MCP_TRANSPORT=http to opt in to Streamable HTTP transport (Bearer auth, /health endpoint).

Categories

Toggle with DBT_TOOLS=dbt (allowlist) or DBT_DISABLE=quality (denylist).

Tools at a glance

dbt (15 + 3)

dbt-list-models, dbt-get-model, dbt-list-tests, dbt-get-test, dbt-list-sources, dbt-get-source, dbt-list-exposures, dbt-list-macros, dbt-get-macro, dbt-list-runs, dbt-get-run-results, dbt-failed-tests, dbt-slow-models, dbt-coverage, dbt-graph, freshness-status, incident-context, dbt-sla-status

quality (6 + 2)

dq-list-checks, dq-get-check-history, dq-failed-checks-by-dataset, dq-score-trend, dq-tier-status, dq-tier-by-source, failed-tests-summary, dq-score-snapshot

Prompts

Environment variables

DQ result-table schema flavors

The quality category supports two schema presets via DQ_SCHEMA:

DQ_SCHEMA=generic (default)

Columns assumed on DQ_RESULTS_TABLE: run_at, check_name, check_type, dataset, table_name, status, severity, failure_count, message.

Columns assumed on DQ_SCORE_TABLE: score_date, scope, tier, completeness_pct, freshness_pct, validity_pct, anomaly_free_pct, overall_score.

dq-tier-status rolls up by Tier 1/2/3 against the per-scope rows.

DQ_SCHEMA=us-all

Real schema used at us-all (Postgres data_ops database):

quality_checks: run_date, check_type, dimension, source, target_name, status, metric_value, threshold, details (JSONB).

quality_score_daily: run_date, completeness_pct, freshness_pct, validity_pct, anomaly_free_pct, overall_score, total_checks, failed_checks.

In this flavor quality_score_daily is one row per day (no per-scope rollup, no tier column). dq-tier-status falls back to comparing the day's overall_score against DQ_TIER1_TARGET_PCT (default 99.5).

dq-get-check-history requires checkName formatted as '<check_type>:<target_name>' since us-all has no native check_name column.

Per-column overrides — DQ_COL_*

If your DQ tables don't match either preset, layer per-column overrides on top of DQ_SCHEMA. Any DQ_COL_* env var, when set, replaces the preset value for that single column. Unset vars keep the preset default.

Overrides are validated as simple SQL identifiers to avoid injecting raw SQL through environment variables. Table names in DQ_RESULTS_TABLE / DQ_SCORE_TABLE are also validated and quoted for the configured backend.

For the three nullable columns (DQ_COL_CHECK_NAME, DQ_COL_SCOPE, DQ_COL_TIER), set the value to none / null / - to declare "no native column":

• Without check_name → the tools synthesize one from check_type || ':' || table_name. dq-get-check-history then expects checkName formatted as '<check_type>:<table_name>'.
• Without scope → dq-score-trend's scope filter is ignored (with a caveat) and dq-tier-status switches to the single-overall_score path that compares against DQ_TIER1_TARGET_PCT.
• Without tier → same single-overall_score fallback.

Example — generic preset against a Postgres schema where columns happen to be named differently:

DQ_SCHEMA=generic
DQ_COL_RUN_AT=checked_at
DQ_COL_DATASET=schema_name
DQ_COL_TABLE_NAME=tbl
DQ_COL_FAILURE_COUNT=fail_n
DQ_COL_CHECK_NAME=none      # synthesize from check_type+tbl
DQ_COL_SCOPE=none           # no per-team rollup
DQ_COL_TIER=none            # use DQ_TIER1_TARGET_PCT instead

SLA config (optional) — DBT_SLA_CONFIG_PATH

Set DBT_SLA_CONFIG_PATH to a YAML file to surface project-defined tier targets and DBT SLAs to the quality tools. Schema (extra keys ignored):

dbt_sla:
  test_pass_pct: 99.0          # consumed by dbt-sla-status (test pass rate threshold)
  freshness_pass_pct: 99.5     # consumed by dbt-sla-status (source freshness pass rate threshold)

tier_sla:
  1: 99.5                      # tier-1 overall_score / per-source pass-rate target
  2: 99.0
  3: 95.0

When set, the tier_sla map drives:

• dq-tier-status — per-tier rollup compares each row's overall_score against the matching target. Without this file, hardcoded {1: 99.5, 2: 99.0, 3: 95.0} is used.
• dq-tier-by-source — per-source pass-rate is compared to the target for that source's tier (resolved from dbt sources.yml meta.tier).
• dq-tier-status no-tier-column path (us-all preset / DQ_COL_TIER=none) — uses tier_sla.1 as the single target. DQ_TIER1_TARGET_PCT env still works as a fallback when no SLA file is set.

The dbt_sla block drives:

• dbt-sla-status — computes test pass rate from latest run_results.json and freshness pass rate from sources.json, then compares each axis against dbt_sla.test_pass_pct / dbt_sla.freshness_pass_pct. Returns passPct, target, meeting per axis plus caveats when fields or artifacts are missing.

The file is mtime-cached; edits between tool calls are picked up automatically.

Per-tier rollup from quality_checks — dq-tier-by-source

For schemas where quality_score_daily has only one row per day (no per-scope/tier breakdown), dq-tier-by-source reconstructs a per-tier picture from the raw quality_checks rows. Two modes:

mode: "source" (default) — group by source/dataset column

Use when each row of quality_checks represents a check on a source group and the dataset/source column carries the dbt source-group name directly.

1. Builds a source_name -> tier map from the dbt manifest's sources.<source>.<table>.meta.tier (first table's tier per source group).
2. Groups quality_checks rows by the dataset/source column and computes pass rate per source over a date or sinceHours window.
3. Looks up each source's tier and target (from SLA config or defaults), reports meeting / missing per tier.

mode: "table" — group by table_name column

Use when the dataset/source column is a category (bq / dbt / airflow) and the actual dbt source-table identifier lives in the table_name / target_name column as <source_group>.<table>. Common in checks tables that consolidate signals from heterogeneous backends.

1. Builds a <source_group>.<table> -> tier map from the manifest using each source entry's source_name + name + meta.tier — picks up table-level tier overrides naturally.
2. Groups quality_checks rows by the table_name column. Pre-filter via sourceFilter (e.g. sourceFilter: "bq") when only some categories produce parseable target names.
3. Each rollup key is parsed as <source_group>.<table>; rows without a . or whose key is not in the manifest land in caveats[].

Untiered rows (no manifest meta.tier) and unparseable rows always appear in caveats[] so you can tier them or accept the gap.

Tested-against schemas

• dbt manifest schema v11 / v12 — the current top version. dbt 1.7 emits v11; dbt 1.8 through 1.12 all emit v12 (the schema evolves additively in-place). Newer/unknown versions still parse, but a caveats line will flag them.

Companion server

For Airflow DAG operations (list, runs, task instances, log tail, trigger, clear), install @us-all/airflow-mcp alongside this server.

Build

pnpm install
pnpm run build      # tsc → dist/
pnpm test           # vitest
pnpm run smoke      # spawns dist/index.js, calls initialize + tools/list (set env first)

License

MIT — see LICENSE.