sqlai.dev SQL Verifier MCP Server

Execute a SQL query against a fresh ephemeral in-memory database built from your schema (and optional seed rows). Returns real rows (max 500, truncation flagged), column names+types, row_count, and dialect notes. Errors come back as structured JSON with type/position/suggestion — a failed query is a useful answer, not a failure of this tool. Example: schema "CREATE TABLE t(id INTEGER, name TEXT);", query "SELECT name FROM t WHERE id=1", seed {"t":[{"id":1,"name":"ada"}]} → rows [["ada"]].

Run up to 10 queries against the same schema+seed. Each query executes in its OWN fresh database — writes in one query are NOT visible to the next (use this for testing variants, not for multi-statement transactions). Returns an array of run_sql results in order.

Answer "do these two queries return the same thing?" — the self-check for query refactors. Executes query_a and query_b against identical fresh databases and compares result multisets (order-insensitive; order divergence reported separately when ORDER BY is present). Returns equal:boolean, row counts, and capped row-level diffs (only_in_a / only_in_b).

Return the engine-native query plan for a query (SQLite: EXPLAIN QUERY PLAN) plus full-table-scan warnings. Use it to check whether an index would be used before recommending one. Example: "SELECT * FROM orders WHERE status=?" on an unindexed column → plan ["SCAN orders"], warning about the full scan.

Validate a SQL query against a schema WITHOUT executing it (parse + name/type binding via EXPLAIN). Returns ok with referenced tables, or a structured error: {type: unknown_column|unknown_table|syntax|..., message, position, suggestion}. The suggestion is rule-based (edit distance against your schema). Example: query "SELECT nmae FROM users" → error type unknown_column, suggestion 'did you mean "name"?'.