Ask your database in plain English. Get correct SQL.

txt2sql runs on your messy, real-world SQL Server — hundreds of cryptic tables, missing foreign keys, no clean docs — and writes correct, production-grade SQL. On-prem or air-gapped. Your data never leaves the building.

Built for teams that can't send their schema, let alone their data, to a cloud API.
The proof

ChatGPT vs txt2sql — same messy schema, same hard question.

▶ Demo video goes here Embed the side-by-side recording (host on YouTube/Vimeo or self-host an .mp4).
Replace this block with the <iframe> / <video> embed.

One question — "customers with an incoming transfer in January but no outgoing transfer in February, dealing with 3+ distinct counterparties." Negation, two date windows, a distinct-count threshold, mixed direction — on a schema with archive and log tables mixed in. A general LLM picks the wrong table, drops the negation, and ships plausible-but-wrong SQL. txt2sql narrows dozens of tables to the few that matter, understands the question, and writes SQL that runs.

Why general LLMs fail on real databases

“Just use ChatGPT” breaks on three things.

Cryptic, undocumented schemas

Hundreds of tables with cryptic names like MatSvcId, relationships that live in someone's head instead of foreign keys, no docs. txt2sql translates cryptic physical names to meaningful ones, infers the relationships that aren't declared, and hides the archive/log/system tables that throw the model off.

Genuinely hard questions

“Who did X but not Y, across two periods, above a threshold?” is where text-to-SQL tools break. txt2sql decomposes the question into a structured plan first — negation, date windows, distinct-count thresholds, source vs destination direction — then generates SQL against that plan.

Data that can't leave the building

For banks, insurers, public sector and healthcare, sending schema or data to a cloud API is a non-starter. txt2sql runs in two modes: cloud mode sends only abstracted schema metadata — never a single row — and local mode sends nothing at all.

The privacy spectrum

Your data stays home. You choose how far that goes.

ModeWhat leaves your network
Local
SQLCoder / Ollama / LM Studio
Nothing   Air-gapped. Even table names stay inside.
Cloud
e.g. Gemini
Schema metadata only   Table/column names, abstracted so noise tables are hidden. Never any row data.

“Row data never leaves, in any mode. If you don't even want schema names to leave, local mode sends nothing at all.”

How it works

Connected to correct SQL in four steps.

  1. Connect. Point it at your SQL Server with a connection string.
  2. It learns your schema — automatically: table roles, descriptions, synonyms (English and Turkish), and the relationships your database never declared as foreign keys. Privacy-safe — only metadata is analyzed, never your rows.
  3. You set the business rules — the things only your team knows (e.g. “CustomerId < 10000 means a corporate account”) get added to your own database, by your admin.
  4. Ask in plain language. It picks the right tables, plans the query, generates validated SQL, and runs it. A validator agent checks every query for syntax, logic and security before it touches your data.
Built for production, not a toy

Safe by default.

The bigger picture

Today it writes your SQL. Tomorrow it runs your data operations.

txt2sql is the entry point. The goal is an autonomous data analyst that lives next to your database — and never sends your data anywhere.

Natural-language SQL Available now

Correct queries on messy, real-world schemas. On-prem.

🩺

Database health Next

Missing indexes, runaway tables, concrete tuning suggestions — read by an LLM that actually explains why.

📈

Scheduled monitoring + interpretation Then

Your recurring checks run on a schedule and the results get interpreted, on-prem — the part cloud tools can't do, because the data can't leave.

🤖

Autonomous operations The goal

Pulls tasks from your tracker, runs the checks, drafts the fix, and routes it to a human for approval. A data ops coworker, fully inside your walls.

Want to shape this roadmap? The first design partners get a direct line.

Who it's for

If your schema can't leave the building, this is for you.

If you run on-prem SQL Server, have privacy or GDPR constraints, and a data team drowning in ad-hoc report requests from the business — txt2sql is built for exactly your situation. Especially if “just use ChatGPT” is off the table because your schema can't leave the building.

See it on your own schema.

A short call, a live demo against a schema like yours, and an honest answer on whether it fits. No deck, no pressure.

Request a demo →

Or email info@txt2sql.app directly.