<!-- https://testapilocally.com/questions/ai-agent-verify-database-migration/ · updated 2026-11-19 -->
# Verifying a database migration with an AI agent

## How can an AI agent verify a database migration?

Give it read-only database tools and have it state each expectation as an assertion. With the SignOff Studio MCP server, the agent:

1. Reads the schema with `db_schema` (one line per table, so it fits in context) and confirms the new columns, types and nullability.
2. Checks the data with `db_assert`. For example, `SELECT count(*) FROM orders WHERE archived_at IS NOT NULL` expects `scalar = 0`, and `SELECT count(*) FROM users WHERE email_normalized IS NULL` expects `0` after a backfill.
3. Checks the API still works on top of it with `verify_api_effect` (call the endpoint, then check the row it should have written).
4. Saves the checks with `save_db_check`, so CI or the next session can rerun them.

The `verify-migration` prompt (a slash command in Claude Code and VS Code) walks the agent through these steps.

## Can the agent break the database while checking?

Not through the read tools. `db_query` and `db_assert` accept only read statements, run them in a read-only transaction (or a read-only SQLite connection, or an always-rolled-back transaction on SQL Server), and run a single statement per call. Writes need `db_execute`, which works only on connections you marked **Allow AI agent writes**, never on production-tagged connections, and runs as a dry run (rolled back) unless the agent asks to commit.

## Which databases are supported?

PostgreSQL, MySQL/MariaDB, SQL Server, Oracle and SQLite.

## How does the agent get the database password?

From the connection's password env key in the workspace `.env`, or from the SignOff Studio app's vault while it's unlocked. The password goes to the driver only. It's never returned to the model, and it's scrubbed from any output.

## Should I run migrations through the agent?

We'd keep running migrations in your usual tool (your framework's migrate command or CI) and use the agent to **verify** the result. Verification is read-only and repeatable. Running migrations isn't.

See also: [Safe read-only SQL access for AI agents](/questions/read-only-sql-access-for-ai-agents/) · [MCP server overview](/mcp/)
