@us-all/dbt-mcp
@us-all/dbt-mcp
dbt MCP 服务器 — 通过一个 stdio MCP 接口提供
manifest.json、run_results.json、sources.json、catalog.json以及 DQ(数据质量)结果表(BigQuery / Postgres)。基于@us-all/mcp-toolkit构建。
为 LLM 客户端提供一个查看 dbt 项目的只读窗口。无需触发 dbt run — 仅提供深度内省、运行历史分析、源新鲜度、列级测试覆盖率、血缘关系遍历,以及(如果您有自定义的 DQ 结果表)历史检查趋势和层级 SLA 状态。
如需 DAG 触发 / 运行历史 / 日志追踪,请同时安装配套的 @us-all/airflow-mcp。
涵盖 3 个类别的 22 个工具 (
dbt,quality,meta)4 个用于分类工作流的 MCP Prompts
4 个聚合工具,可替代 3-5 次“列出 / 获取 / 列出”的往返调用
默认只读
混合后端:BigQuery(默认)或 Postgres 用于 DQ 结果表 — 两者均按需导入
安装
# 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 pgRelated MCP server: dbt-mcp
运行
DBT_PROJECT_DIR=/path/to/dbt-project \
DQ_RESULTS_TABLE=my-project.data_ops.quality_checks \
npx @us-all/dbt-mcp该服务器使用 MCP stdio 协议;将其连接到 Claude Desktop / Cursor / 任何 MCP 客户端。设置 MCP_TRANSPORT=http 可选择使用 Streamable HTTP 传输(Bearer 认证,/health 端点)。
类别
类别 | 工具 | 用途 |
| 15 + 2 个聚合工具 | 解析 |
| 5 + 2 个聚合工具 | 查询 |
| 1 (始终开启) | 用于自然语言工具发现的 |
使用 DBT_TOOLS=dbt(允许列表)或 DBT_DISABLE=quality(禁用列表)进行切换。
工具概览
dbt (15 + 2)
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
quality (5 + 2)
dq-list-checks, dq-get-check-history, dq-failed-checks-by-dataset, dq-score-trend, dq-tier-status, failed-tests-summary, dq-score-snapshot
Prompts
Prompt | 使用场景 | |
| “过去 24 小时内有什么故障?” | |
| “有数据源过期了吗?”(可选关注 Tier 1) | |
| “给我一份利益相关者友好的 DQ 趋势报告” | |
| “分类 <模型 | 源>” — 汇总所有信号 |
环境变量
环境变量 | 必需 | 说明 |
| 是 | dbt 项目根目录 ( |
| 否 | 默认为 |
| 否 | 用于存档 |
| 否 |
|
| 否 | 检查表的 FQN(没有它, |
| 否 | 每日评分表的 FQN |
| 否 | 用于 BigQuery 后端(支持 ADC 回退) |
| 否 | 显式 BQ 项目(否则从 ADC 推断) |
| 否 | 当 |
| 否 | 为未来的写入工具预留(v0.1 中暂无) |
| 否 | 类别切换 |
DQ 结果表假设架构 (v0.1)
quality 类别假设 DQ_RESULTS_TABLE 包含 check_name, check_type, dataset, table_name, status, severity, failure_count, run_at, message 列;假设 DQ_SCORE_TABLE 包含 score_date, scope, tier, completeness_pct, freshness_pct, validity_pct, anomaly_free_pct, overall_score 列。v0.2 将增加可配置的列映射层。
已测试架构
dbt manifest 架构 v11 / v12 / v13(其他版本通常可以解析,但会显示
caveats行进行提示)
配套服务器
如需 Airflow DAG 操作(列出、运行、任务实例、日志追踪、触发、清除),请与此服务器一起安装 @us-all/airflow-mcp。
构建
pnpm install
pnpm run build # tsc → dist/
pnpm test # vitest
pnpm run smoke # spawns dist/index.js, calls initialize + tools/list (set env first)许可证
MIT — 参见 LICENSE。
Available Tools
27 toolsdbt-coverageA
Per-column test coverage for a dbt model (which columns have tests, table-level tests, coverage %)
Inputs: name, uniqueId, extractFields.
dbt-failed-testsB
Find tests that failed across the last N runs, grouped and ordered by chronic failure count
Inputs: recentRuns, extractFields.
dbt-get-macroA
Get a dbt macro: signature, raw SQL, and reverse-lookup of nodes that call it
Inputs: name, uniqueId, extractFields.
dbt-get-modelA
Get a single dbt model: refs, sources, columns (with catalog types if available), attached tests, raw/compiled SQL
Inputs: name, uniqueId, extractFields, includeCompiledSql.
dbt-get-run-resultsB
Get per-node results from a specific dbt invocation (or the latest run if invocationId omitted)
Inputs: limit, status, invocationId, extractFields.
dbt-get-sourceA
Get a single dbt source: freshness criteria, columns, latest freshness result from sources.json
Inputs: uniqueId, tableName, sourceName, extractFields.
dbt-get-testA
Get a single dbt test: definition, parameters, attached models, latest run result
Inputs: uniqueId (required), extractFields.
dbt-graphB
Walk dbt parent_map / child_map to return upstream and downstream nodes (model/source/test) up to a given depth
Inputs: name, uniqueId, extractFields, upstreamDepth, +1 more.
dbt-list-exposuresA
List dbt exposures (downstream BI/ML/application consumers declared in YAML)
Inputs: limit, search, exposureType, extractFields.
dbt-list-macrosA
List dbt macros from manifest.json with package / name filters
Inputs: limit, search, package, extractFields.
dbt-list-modelsA
List dbt models from manifest.json with filters (package, tag, materialized, schema, name search)
Inputs: tag, limit, schema, search, +3 more.
dbt-list-runsC
List recent dbt invocations from run_results.json files in target/ and DBT_RUN_HISTORY_DIR
Inputs: limit, extractFields.
dbt-list-sourcesC
List dbt sources from manifest.json with optional source-group / name filters
Inputs: limit, search, sourceName, extractFields.
dbt-list-testsC
List dbt tests (generic or singular) optionally filtered to a specific model
Inputs: limit, search, testKind, attachedTo, +1 more.
dbt-sla-statusB
Compare latest dbt run test pass rate and source freshness pass rate against DBT_SLA_CONFIG_PATH thresholds (dbt_sla.test_pass_pct / freshness_pass_pct). Returns per-axis passPct/target/meeting plus caveats when SLA fields or artifacts are missing.
Inputs: extractFields.
dbt-slow-modelsB
Top N slowest models in a dbt run by execution_time, with bytes_processed when available
Inputs: topN, invocationId, extractFields.
dq-failed-checks-by-datasetB
Group failing checks by dataset across a recent window with the latest 5 failures per dataset
Inputs: topN, sinceHours, extractFields.
dq-get-check-historyC
Time-series of one check_name's status across the last N days
Inputs: checkName (required), days, limit, extractFields.
dq-list-checksA
List recent rows from DQ_RESULTS_TABLE filtered by dataset / status / type / time window
Inputs: type, limit, status, dataset, +2 more.
dq-score-snapshotA
Aggregated 4-axis score trend + today's Tier compliance + most recent failing checks. Combines dq-score-trend + dq-tier-status + dq-list-checks(fail).
Inputs: days, extractFields, includeFailing.
dq-score-trendB
Time-series of the 4-axis DQ score (completeness / freshness / validity / anomaly_free) plus overall_score from DQ_SCORE_TABLE
Inputs: days, scope, extractFields.
dq-tier-by-sourceA
Per-tier rollup computed from quality_checks grouped by source. Reads source-to-tier mapping from dbt sources.yml meta.tier and tier targets from DBT_SLA_CONFIG_PATH (falls back to defaults). Reports per-source pass rate, meeting/missing per tier, and untiered sources as caveats.
Inputs: date, mode, sinceHours, sourceFilter, +1 more.
dq-tier-statusA
Compare today's overall_score per scope against Tier SLA targets (defaults Tier 1 99.5 / 2 99.0 / 3 95.0; override via DBT_SLA_CONFIG_PATH yaml or DQ_TIER1_TARGET_PCT) and report meeting vs missing counts
Inputs: date, extractFields, fallbackMaxDays.
failed-tests-summaryA
Aggregated 24h-ish view: dbt failed tests + DQ checks failures grouped by dataset + most recent failing rows. Replaces 3+ tool calls (dbt-failed-tests + dq-failed-checks-by-dataset + dq-list-checks).
Inputs: recentRuns, sinceHours, extractFields.
freshness-statusB
Cross-reference dbt source freshness criteria with sources.json results in a single 'is anything stale right now?' answer.
Inputs: failingOnly, extractFields.
incident-contextA
Single asset deep-dive: dbt definition + recent test failures + DQ checks for the dataset. Designed to anchor an LLM-driven incident triage.
Inputs: modelName, sourceFqn, sinceHours, extractFields.
search-toolsA
Discover available dbt MCP tools by natural language query across dbt / quality / meta categories.
Inputs: query (required), limit, category.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 27 tools
Most tools have distinct purposes, but composite tools like `failed-tests-summary` overlap with `dbt-failed-tests` and `dq-failed-checks-by-dataset`, and `incident-context` aggregates multiple individual tools, potentially causing confusion about which to use.
Tools follow a consistent prefix pattern (`dbt-` and `dq-`) but there are deviations like `dbt-graph` instead of `dbt-get-graph`, and four tools without any prefix (`failed-tests-summary`, `freshness-status`, `incident-context`, `search-tools`), mixing naming conventions.
With 27 tools covering dbt core operations, data quality, and incident triage, the count is slightly high but well-justified for the broad scope, avoiding the extremes of overwhelming or insufficient coverage.
The tool set covers essential dbt monitoring and data quality workflows (listing, details, failures, SLAs, scores), but lacks write operations and some administrative functions, though that seems intentional for a read-only analysis server.
Maintenance
Related MCP Connectors
The Cortex MCP server provides read-only access to real-time engineering context from the Cortex developer portal, allowing AI coding assistants to answer natural language questions about your organization's catalog (microservices, libraries, domains, teams, infrastructure), scorecards (engineering standards and best practices), initiatives (goals and deadlines), and Engineering Intelligence metrics. It includes tools for querying documentation, tracking personal entities, and accessing AI-assisted insights across the entire Cortex ecosystem.
Guarded MCP server for agent-readable business truth, provenance, readiness, and discovery.
Query your org's data in natural language — read-only MCP access to SQL, NoSQL, files & warehouses.
Query your warehouse or a CSV with Claude/ChatGPT over MCP, governed by table-level ACL + audit.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceAI-driven MCP server that audits, profiles, detects schema drift, and auto-generates documentation for dbt projects, enabling natural language interaction with your dbt project's health.105MIT
- AlicenseNot gradedqualityDmaintenanceMCP server that provides tools to interact with dbt, including dbt Core, Cloud CLI, Semantic Layer, and Discovery API.Apache 2.0
- AlicenseNot gradedqualityDmaintenanceA production-ready MCP server that provides comprehensive dbt project quality assessment for any GitHub repository, enabling AI agents to analyze dbt models, check metadata coverage, and map data lineage.8MIT
- AlicenseNot gradedqualityDmaintenanceMCP server providing read-only Snowflake metadata tools (schemas, tables, queries, lineage) for agentic data pipeline generation, enabling natural-language-to-pipeline workflows with dbt, Airflow, and Great Expectations.1MIT