TL;DR
Data engineers spend half their day writing boilerplate SQL, the other half explaining data to stakeholders. MCPs flip that: Claude writes the SQL from natural language, inspects the warehouse schema, runs profiling, and generates the ad-hoc dashboard. Postgres, ClickHouse, DuckDB, and S3 as MCPs cover 90% of a data engineer's daily work.
The MCP stack for modern data platforms
A data engineer building pipelines across Postgres, a warehouse (ClickHouse/Snowflake/BigQuery), and object storage. Owns dbt, Airflow, and the BI layer.
What hurts today
- 1Stakeholder Slack pings ('quick question: how many active users last month?') break every single afternoon
- 2New analyst onboarding takes 2 weeks because the schema isn't documented and tribal knowledge is in Slack DMs
- 3dbt test failures at 3am — debugging the exact row that violated a constraint eats the first hour
- 4Building a one-off analysis dashboard is faster to do in DuckDB than in Looker, but context-switching to a notebook kills momentum
- 5Cost attribution on the warehouse is a quarterly fire drill — no day-to-day visibility on who runs what
Recommended MCPs (7)
Postgres
View MCPThe universal data MCP — works with any Postgres-compatible DB (Postgres, Supabase, Neon, Aurora). Schema introspection, query execution, EXPLAIN analysis, all in the IDE.
ClickHouse
View MCPFor the warehouse side — ClickHouse MCP makes stakeholder questions answerable in seconds instead of a PR to the dbt repo.
DuckDB
View MCPLocal analytics on parquet/CSV files. Ideal for ad-hoc analysis, profiling a new data source, or prototyping a pipeline before committing to dbt.
AWS (S3)
View MCPList, inspect, and read from S3 buckets — your data lake lives here. Combined with DuckDB MCP, you can query parquet files on S3 directly from Claude Code.
GitHub
View MCPdbt models, Airflow DAGs, and Dagster assets live in GitHub. Claude reads model definitions, runs dbt test, and opens PRs for schema changes.
Grafana
View MCPDashboards-as-code for monitoring pipeline freshness and row counts. Claude creates dashboards from a one-line description.
Neo4j
View MCPWhen your data model is a graph (lineage, fraud detection, social). Neo4j MCP lets Claude write Cypher queries from natural language.
A real workflow
A PM pings you: 'what's the weekly active user growth over the last 8 weeks, broken down by plan tier?' Instead of opening a BI tool, you prompt Claude in your terminal. It introspects the Postgres schema via the MCP, writes the CTE query, runs it, pivots the result, generates a quick ASCII chart, writes the SQL back into your dbt repo as a new model `weekly_active_by_tier.sql`, opens a PR, and posts the chart+query into Slack for the PM. Total: 8 minutes, zero tab switches.
Time ROI
Data engineers save 10–14h/week. The biggest win is ad-hoc stakeholder questions (−80% time), followed by pipeline debugging and schema documentation.
Recommended recipes for this role
Analytics Dashboard
Connect ClickHouse to Grafana to build real-time analytics dashboards over billions of events with sub-second query times.
Data Lake Queries
Query Parquet files directly from S3 using DuckDB without any ETL. Results are returned in seconds for ad-hoc analytics.
Database Monitoring Dashboard
Stream Postgres metrics — query latency, lock waits, vacuum stats — into Grafana for a live operations dashboard.
Full-Text Search Sync
Keep Meilisearch in sync with your Supabase tables. Inserts, updates, and deletes are reflected in the search index in real time.
Knowledge Graph from Code
Parse your GitHub repos and build a Neo4j knowledge graph of files, functions, imports, and authors for code intelligence.
Frequently asked questions
Can Claude write dbt models for me?
Yes. With the GitHub MCP and a Postgres/ClickHouse MCP for schema introspection, Claude writes dbt models that follow your project's conventions. It reads your existing models for style, generates the new one, runs `dbt test`, and opens a PR.
Is it safe to let Claude run SQL on production Postgres?
Use a read-only replica for most queries. The Postgres MCP supports multiple connections — point it at your replica for exploration, at the primary only for controlled writes behind a confirmation.
What about Snowflake or BigQuery?
Both have community-maintained MCPs. Snowflake MCP supports warehouse/role selection, BigQuery MCP handles project/dataset scoping. Same pattern as ClickHouse MCP.
How does DuckDB MCP help if I already have a warehouse?
DuckDB is your fast local scratch pad. Dump a parquet to /tmp, query it in Claude, validate the transformation, then lift it to dbt. It's 10x faster for prototyping than doing everything in the warehouse.
Can Claude help me document our schema?
Yes — point Claude at your Postgres/ClickHouse MCP and ask 'generate dbt YAML for every table in schema analytics with column descriptions'. Claude reads the schema, infers descriptions from table/column names and sample rows, and writes the YAML.
Other use cases
MCPs for SaaS Founders
A technical founder (0–10 employees) building a B2B SaaS who ships code, handles billing, writes marketing, and answers support — all in the same day.
MCPs for Solopreneurs & Indie Hackers
An indie hacker with a Twitter audience, a newsletter, 1–3 shipped products, and zero employees. Ships daily, markets constantly, avoids meetings.
MCPs for AI Agent Developers
A developer building AI agents, chatbots, or autonomous workflows. Needs search, scraping, vector storage, and LLM orchestration — all as tools the agent can call.
Start with this MCP stack
Install the full stack in one command, or cherry-pick the MCPs you need.
Browse all MCPs