New text-to-SQL repositories are shipping code before they establish release readiness

Four repositories, four different maturity signals

A directory sweep on October 14 found four recently created text-to-SQL repositories that are useful precisely because they expose the difference between visible code and a dependable release.

SQLStage is the most complete workbench of the group: its repository describes configurable schema retrieval, few-shot prompting, evaluation and budgeted grid search across SQLite, MySQL and PostgreSQL. GitHub identifies the project as Apache-2.0 licensed. But the repository still has no tagged release, and its public adoption signal is one star and no forks.

AskData AI packages a FastAPI and React assistant with local embeddings, bring-your-own model credentials and SQLGlot guardrails. Its MIT license is clear, but its repository has not been pushed since its initial September 18 upload. It has no tagged release, stars or forks.

The MIT-licensed nl2sql-eval tackles a more valuable question than another chat interface: whether a Databricks text-to-SQL system should answer, clarify or refuse. That makes it a useful evaluation artifact, but not yet a broadly validated benchmark. GitHub shows no push after October 1, no tagged release, and no stars or forks.

The licensing gap is not cosmetic

SQL AI Harness is the freshest and most instructive case. Its README documents a MySQL CLI agent with table discovery, relationship inspection, read-only SQL, a three-attempt execution cap and optional business context. It also states an important limitation: discovering foreign keys is not an AST-level proof that every generated join follows a declared relationship.

The repository calls itself open source, but GitHub reports no recognized license and the root directory contains no license file. That means readers can inspect the code, but the project has not yet granted the standard reuse rights that an open-source directory entry requires. The project also has no tagged release; GitHub showed two stars and no forks during this audit.

What maintainers need to add

The pattern is not that these projects lack ideas. They cover retrieval tuning, refusal testing, local deployment and join safety. The missing layer is release evidence: a versioned tag, an explicit license, a reproducible evaluation and at least one independent user or rerun.

For directory readers, those signals should remain separate from feature claims. A README can establish intended behavior. It cannot establish stability, legal reuse, benchmark performance or adoption. Until those artifacts arrive, the right label is early implementation—not production-ready tool.