Ask Claude or ChatGPT about your warehouse data
Connect Claude or ChatGPT to the warehouse Ingest loads through that warehouse's own MCP server, with a read-only login, and Ingest out of the path.
Use the MCP server your warehouse already provides. Snowflake, BigQuery, Databricks, ClickHouse Cloud and MotherDuck host one; for Redshift, Postgres, MySQL, Athena and files in your own bucket, an open-source server runs on your computer. You give it a login that can only read, add it to Claude or ChatGPT as a connector, and ask questions in plain words. Your AI app talks to your warehouse directly: Ingest is not in the path and sees no question or answer.
Which MCP server does each warehouse use?
| Where Ingest lands your data | MCP server | Works in | |---|---|---| | Snowflake | A Snowflake-managed MCP server, created in your account with SQL | Claude or ChatGPT | | BigQuery | Google's managed BigQuery MCP server | Claude or ChatGPT | | Databricks | Databricks' managed SQL MCP server, in your workspace | Claude or ChatGPT | | ClickHouse | ClickHouse Cloud's hosted server, or ClickHouse's open-source server on your computer | Claude or ChatGPT | | MotherDuck | MotherDuck's own hosted MCP server | Claude or ChatGPT | | Redshift | AWS's open-source Redshift MCP server, through the Redshift Data API | Claude Desktop, Claude Code | | Postgres, MySQL | Google's open-source MCP Toolbox for Databases | Claude Desktop, Claude Code | | Athena | AWS's open-source Data Processing MCP server, against the Iceberg tables in Glue | Claude Desktop, Claude Code | | S3, Google Cloud Storage, Azure, Iceberg | DuckDB, through MotherDuck's open-source MCP server, reading the files Ingest writes | Claude Desktop, Claude Code |
A server that runs on your computer works in Claude Desktop and Claude Code, not in Claude or ChatGPT on the web. If your Postgres runs on a provider that hosts its own MCP server, such as Supabase, Neon, Google Cloud SQL or AlloyDB, the web apps can reach that one instead.
How do I keep it read-only?
Give the MCP server its own login, one that can read the schemas Ingest writes to and nothing else, and never reuse the login Ingest writes with. Then make the server refuse writes where it can:
- On Snowflake, the server's only tool runs SQL with
read_only: true, so it refuses anything but a read even for a role that could write. - On BigQuery, block the
execute_sqltool in Claude's connector settings. ChatGPT asks before any call that is not marked read-only; decline it there.
Queries run, and bill, in your own warehouse, on the compute you name.
Where is the step-by-step setup?
In Ingest, open MCP in your workspace. It walks through the setup for the destination your pipelines use, with your own database, schemas and warehouse already filled into every command.
What should I ask first?
Start with questions whose answers you can check:
- "What tables are in this schema, and how many rows does each have?"
- "When was each table last loaded?" Ingest records every load in
_ingest_loads, with its status and row count (The tables Ingest creates).
Before you rely on an answer, ask which tables and filters it used.
Why can an answer from raw tables be wrong?
The tables Ingest lands are shaped like the source, and a question can be right about the SQL and wrong about the business. A list inside a record, such as an order's line items, lands as its own child table, so a join from orders to line items repeats each order once per item, and a total taken after that join counts the order's amount more than once. Curated tables exist for this: on Snowflake, Ingest can design silver and gold tables whose grain and joins are checked before they are built and every day after (How to check the tables Ingest lands). Point your questions at those where you have them.
For when to query a vendor's API live instead of a warehouse, see Direct MCP vs Warehouse.
Where this applies
Sources certified against a real API today: Abilene TX 311, Albany NY 311, Ann Arbor MI 311, Anne Arundel County MD 311, Anthropic Admin API (Claude API), Asana, Auburn WA 311, Aurora IL 311, Austin TX 311, Baltimore MD 311, Baton Rouge LA 311, Beaumont TX 311, Bellevue WA 311, Bellingham WA 311, Berkeley CA 311, Bethlehem PA 311, Bloomington IN 311, Boca Raton FL 311, Boston MA 311, Bridgeport CT 311, Brockton MA 311, Brookline MA 311, Buffalo NY 311, Burlington VT 311, Cambridge MA 311, Cambridge MA Open Data, Charleston County SC 311, Charlotte NC 311, OpenAI (ChatGPT Enterprise), Chula Vista CA 311, Cincinnati OH 311, Chicago IL 311, New York City NY 311, Clayton County GA 311, Cleveland OH 311, Clockify, Cobb County GA 311, Columbus OH 311, Corona CA 311, Azure Cosmos DB, Dallas TX 311, Davenport IA 311, Denver CO 311, Detroit MI 311, Elgin IL 311, Encinitas CA 311, Everett WA 311, Federal Way WA 311, Fort Lauderdale FL 311, Gilbert AZ 311, Google Ad Manager, Google Analytics 4, Google Drive, Google Sheets, Greenwich CT 311, Hartford CT 311, Honolulu HI 311, Houston TX 311, Howard County MD 311, Huntington WV 311, Huntsville AL 311, Indianapolis IN 311, Jersey City NJ 311, Jira, Kansas City MO 311, Kent WA 311, Knox County TN 311, Los Angeles CA 311, Lexington KY 311, Lincoln NE 311, LinkedIn Ads, Livonia MI 311, Louisville KY 311, Lubbock TX 311, Macon-Bibb County GA 311, Manchester NH 311, Maui County HI 311, Medford MA 311, Memphis TN 311, Miami FL 311, Miami-Dade County FL 311, Minneapolis MN 311, Montgomery County MD 311, Multnomah County OR 311, Naperville IL 311, Nashua NH 311, Nashville TN 311, Newark NJ 311, NOAA, New Orleans LA 311, Norfolk VA 311, Oakland CA 311, Omaha NE 311, OpenAI Admin API (API Platform), Open Brewery DB, Paterson NJ 311, Peoria IL 311, Philadelphia PA 311, Pinellas County FL 311, Pittsburgh PA 311, PokéAPI, Portland ME 311, PostgreSQL, PrimeGov public meetings, Prince George's County MD 311, Provo UT 311, Raleigh NC 311, Ramp, Renton WA 311, The Rick and Morty API, Rippling, Sacramento CA 311, Salesforce, San Antonio TX 311, San Francisco CA 311, San Gabriel CA 311, San Jose CA 311, San Mateo County CA 311, San Rafael CA 311, Santa Fe NM 311, Scottsdale AZ 311, Seattle WA 311, Shopify, Somerville MA 311, Springfield MO 311, St. Petersburg FL 311, Tacoma WA 311, Toledo OH 311, Topeka KS 311, Vallejo CA 311, Washington DC 311, Wayne County MI 311, West Des Moines IA 311, Wichita KS 311, Worcester MA 311, Youngstown OH 311.
Where the tables land: Amazon S3, Google Cloud Storage, Azure Blob Storage, Apache Iceberg, Ingest Managed Lakehouse, MotherDuck, Postgres, Amazon Athena, Databricks, Redshift, BigQuery, ClickHouse, ObsessionDB, MySQL / generic SQL, Snowflake.