NLQueries turns plain-English questions into SQL, checks that SQL against your real schema before it runs, executes it read-only with a timeout, and shows the SQL beside the answer. It learns from your team's real query history, reads your documents (PDF, Word, Excel, Notion, Confluence) so one answer can draw on both, and works inside Claude Desktop and Cursor as an MCP server. Self-hosted, with the LLM you choose. 8 databases plus any SQLAlchemy URL. Free, with the source on GitHub.

Hi Product Hunt 👋 I'm Pankaj, and I built NLQueries.

The problem. Text-to-SQL demos look great until they meet a real database. The model guesses a column that doesn't exist, joins on the wrong key, or writes something that only fails once it's running against production. And the people asking questions can't tell a good query from a bad one.

What NLQueries does differently:

SQL is validated before it runs. Every generated query is parsed and checked against your live schema before it reaches the database, then run read-only with a timeout. Use ask to preview the SQL without touching the database, or query to run it. You always see the exact SQL.
It learns from your real query history. The knowledge base isn't just a schema dump. It's built from the queries your team already runs, so it learns the joins and the naming you actually use.
Databases and documents in one answer. Ingest PDFs, Word, Excel, Notion or Confluence, and ask a question that spans both.
It's a tool, not another chat window. There's a full MCP server, so Claude Desktop, Cursor and other MCP clients can query your database as a native tool, with token authentication and per-tool grants for network clients.

It's self-hosted. Your database credentials stay with you, and you choose the LLM: Anthropic, OpenAI, Amazon Bedrock, DeepSeek, or others through LiteLLM. It connects to PostgreSQL, MySQL, Snowflake, BigQuery, Redshift, SQL Server and Azure SQL, DuckDB and SQLite, plus anything with a SQLAlchemy URL.

On the licence, since someone will ask: Core is under the Business Source License 1.1. The source is public and you can run it in production. The one thing you can't do is sell it to others as a hosted service. Each version converts to Apache 2.0 four years after release, and the whole project on June 4, 2030.

Try it:

pip install nlqueries-core
nlqueries connect postgres --host localhost --database mydb --user alice --alias prod
nlqueries process-history prod --annotate && nlqueries export-kb prod
nlqueries query prod "top 10 customers by revenue this year"

Or run the whole stack with Docker. The quickstart is at https://nlqueries.com/docs/getting-started, and the source is at https://github.com/nlqueries/nlqueries.

I'd love to hear where it gets a question wrong. That's the most useful feedback there is. I'm here all day to answer anything.