Your assistant can already write SQL. What it can’t do on its own is see your data. It guesses from column names and whatever you pasted into the chat. TestGen’s MCP server closes that gap: the model reads a real profile, real test results, and real history, and it cites the run instead of inventing a plausible one.
What the 96 tools cover
The MCP cheat sheet puts all 96 on one page, grouped by stage. Every tool that writes to TestGen carries an orange dot on the sheet, so your security team can see the read and write boundary before anyone clicks “always allow.”
| Stage | An example of what to ask (illustrative) |
|---|---|
| Discover | What data do I have, and how healthy is it? |
| Profile | Profile this table group and show me what changed since the last run |
| Hygiene | What’s dragging the score? Show me the rows behind that issue |
| Test | Generate tests for this table; add a rule that fares are never negative |
| Run | Run the suite and list what failed |
| Monitor | Which tables are late, short, or reshaped? |
| Score | Build a scorecard for these critical data elements |
| Operate | Dismiss the recency issue; it’s historical. Confirm the rest |
The Discover, Hygiene, and Operate prompts come from a real session the cheat sheet describes. In it, one question returned an inventory and a quality score of 78.9. The follow-up surfaced 30 hygiene issues nobody had tested for, including a “clean” table still carrying names, emails, zip codes, and income flagged as PII. “Generate tests and run them” came back with 12 failures that matched the hygiene story, and every disposition landed in TestGen with an audit trail.
Which clients, and what “supported” means here
MCP is a standard, so TestGen doesn’t ship one integration per assistant. The product page lists Claude, Claude Code, Cursor, GitHub Copilot, and Databricks Genie. DataKitchen’s published walkthroughs use Claude and Claude Code. The NYC taxi run had Claude Code read a TestGen profile over MCP and write seven business rules. The cheat sheet’s setup steps don’t name a client: point it at the /mcp path on your TestGen host, sign in through your browser, approve the connection, and ask it to list your projects. DataKitchen hasn’t published a Cursor or Copilot walkthrough, so check that your client supports remote MCP servers and read the setup guide.
The server authenticates as you, and every tool runs with your existing TestGen role, so the assistant can do what you can and nothing more. An analyst’s assistant reads scores. It can’t delete scorecards.
Triage first, remediation with a gate
DataKitchen’s advice in While You Slept, an Agent Fixed 14 Data Quality Failures is to start with triage. Let the model explain failures and recommend tests before you give it anything to change. You learn how good its reasoning is on your data at no risk.
When you do let it fix things, keep a human in the loop. The design pattern DataKitchen published in June 2026 splits the job in two. Program one uses TestGen and Claude to find problems and propose exact SQL, filed as a Jira ticket with the affected row counts. A person approves or rejects each fix. Program two applies only the approved ones, in one transaction. The agent writes the SQL well. It can’t decide whether 222 invalid state codes should become NULL or map to real states.
That’s a business call, and it stays yours.
Where the dashboard still wins
Not every job belongs in a chat window. The Stop Clicking, Start Asking webinar argues for three ways of working (the UI, a dialogue, and delegation to an agent) and says the UI still wins for some jobs. Scanning a scorecard you read every morning is faster than asking for it.
An assistant also can’t grade its own homework. It writes the tests and reads the results, and TestGen runs them against real rows. If you want to know what the model writes well and where it stops, can AI write your data quality tests has the numbers from two public datasets.