Claude Dashboards on Wren AI: One Question, 24 Governed Queries

Claude can now turn one prompt into a full dashboard. That changes the question a data team has to answer. It is no longer "can the model write SQL?" It is "would I put this dashboard in front of the CEO?"

The answer depends on what Claude is connected to. Pointed at a raw schema, it guesses what columns mean, picks its own joins, and runs with whatever access a shared service account has. Connected through Wren AI, it reads your context layer before it writes a single query, and every number it puts on the page traces back to governed SQL.

Here is the whole flow in 44 seconds.

20+ databases, one connector

Most companies do not keep their data in one place. Sales sit in Postgres, finance in Snowflake, product events in BigQuery, and something important is still in SQL Server.

Wren AI gives Claude one MCP connection that reaches every source Wren AI supports: PostgreSQL, SQL Server, Oracle, MySQL, Snowflake, BigQuery, Databricks, Redshift, ClickHouse, Athena, Trino, Starburst and more, 20+ in all. There is no ETL step and no migration. The Wren AI engine plans each query against the source where the data already lives.

For Claude, that means a dashboard question does not stop at the edge of one warehouse.

A context layer Claude reads first

Before Claude writes any SQL, it reads the model. In the demo, that is a retail project built around a Store Master table, connected to monthly POS sales, store KPIs, human resources, supply chain inventory, hourly sales, member RFM segments and new-store ramp-up.

The context layer tells Claude the things a column name never will:

• What a column means. floor_area_ping is store floor area measured in ping, where one ping is about 3.3 square meters. Without that, "revenue per square meter" comes out wrong by a factor of three.
• How tables join. Store Monthly KPI joins to Human Resources on store_id and year_month, not on store_id alone. The relationship is declared, so Claude does not have to discover it, and every caller gets the same join.

These definitions live in MDL, Wren AI's context model, as git-versioned files in your own repository. Change a definition in a pull request and Claude uses the new one on its next question.

One question, 24 queries, one thread

The prompt in the demo is one sentence: using the Mart dataset in Wren AI, build an executive dashboard for the CEO and CIO on the store expansion plan, clear, visual and boardroom-ready.

Claude breaks that into 24 queries and runs them through Wren AI's run_sql tool: member RFM segments, staffing and supply chain, revenue by trade area, format and city, ramp-up by month since opening, store-level positions, and more.

The interesting part is between the queries. Partway through, Claude notices that hard-coded months go stale and switches to a derived rolling 12-month window, then runs a query to test that window before relying on it. Later it goes back and verifies its numbers: format economics, network by month, trade area and format, store positions. Up to five queries run in parallel, and ten of them are kept as live sources behind the final dashboard.

That is what a context layer buys you. Claude spends its reasoning on the analysis, not on guessing what the data means.

Then it builds the dashboard

The result is a Store Expansion Outlook that reads like something an analyst would bring to the board. It opens with a short narrative: transit-hub stores earn 25.2% more per month than the average store, a new store reaches full maturity in month 10, and the network is growing 3.5% year to date, so new doors should go where traffic is densest, in compact, high-yield formats.

Below that sit the headline KPIs, a four-part case for where, what, how fast and which cities to expand, revenue per store by trade area, and a trade area by store format heat map.

The demo's numbers come from a sample retail dataset. What carries over to your data is the structure: the dashboard runs on 10 live queries, and every number on it links to its SQL. When someone in the meeting asks "where does 153.1K come from?", the answer is one click away. It is not buried in a chat transcript.

Governed, not just generated

A dashboard built by an AI model is only as trustworthy as the path the data took to get there. With Wren AI in that path:

Access rules apply at query time to the person asking, so two people asking for the same dashboard see only the rows and columns each is allowed to see. Every request is logged with who asked and the SQL that ran.

Try it

1. Add Wren AI to Claude. Search "Wren AI" in Claude's connector directory and authorize it with OAuth. The connector announcement covers setup in detail.
2. Run the open-source engine. Wren AI core is open source on GitHub. You can run the whole path locally against your own database first.
3. Govern it at scale. Row- and column-level security and audit logs are part of Wren AI's paid plans.

Claude builds the dashboard. Wren AI makes sure every number on it means what your business means.