Postgres.new: In-browser Postgres with an AI interface

Original Article ↗ Hacker News Discussion ↗

Project overview & components

  • postgres.new is an in-browser Postgres “sandbox” built on a WASM Postgres (PGlite), with a chat-style interface that generates and runs SQL.
  • Frontend, PGlite, pg-gateway, and transformers.js usage are all open source; only the LLM service is not.
  • Users praise the tool as impressive, fun, and surprisingly capable at designing schemas, constraints, and example data.

AI integration & model choice

  • Current version is tightly coupled to GPT‑4o; GitHub login is required “to prevent abuse,” which in practice gates the whole app, not just chat.
  • Some users like the AI workflow and report high accuracy for complex SQL; others argue alternative or specialized models benchmark better.
  • There is strong demand for:
    • A fully local / offline LLM option.
    • A mode that keeps the UI and visualizations but removes mandatory AI.
    • A way to send raw SQL without going through the model.

UX, access & platform support

  • Multiple complaints that the UI doesn’t clearly communicate that login and AI are required even to create a database.
  • Confusion around the “New database” button and the need to “start typing” after clicking.
  • Mobile is currently blocked; on desktop, small windows and missing browser APIs (OPFS, IndexedDB differences) cause “use a laptop/desktop” warnings.
  • Safari support is somewhat unclear: some users see blocking, others report it works.

Use cases & capabilities

  • Suggested uses: playgrounds, teaching SQL, quick schema prototyping, offline/local data store, deploying playgrounds to the cloud for fast project bootstrapping.
  • Users want shareable databases and ER diagrams, including CI/CD-generated diagrams from existing schemas.

Skepticism & limitations

  • Several users dislike “LLMs in my DB,” preferring a classic playground without AI.
  • Concerns include:
    • Invalid or subtly wrong SQL requiring careful review.
    • Overreliance on AI eroding database and query design skills.
    • “Good enough” schemas and data models becoming widespread.
  • Others counter that LLMs already save substantial time on SQL and code, especially for non-experts, even if outputs must be reviewed.