Show HN: Refine – A Local Alternative to Grammarly

Original Article ↗ Hacker News Discussion ↗

Privacy & Local Processing

  • Many see the main differentiator vs Grammarly as being fully local processing and no cloud data transfer, especially for corporate/IT environments.
  • Some argue BYOK (remote models) risks diluting that advantage; others want BYOK to run heavier models on home servers.
  • Several suggest the marketing should strongly emphasize privacy and show a clear comparison with Grammarly and Apple’s system tools.

Language Support & Dialects

  • Underlying model (Gemma 3n) can theoretically support ~140+ languages; real-world quality beyond English is largely untested.
  • Big concern that the site doesn’t clearly specify supported languages, dialects, or registers (e.g., US vs UK English, Indian English).
  • Debate erupts over “practice/practise” and spellings like “colour,” with some saying correct dialect handling is a minimum requirement for a serious checker.

Quality & Behavior of Suggestions

  • Users find the “fluency” mode often over-aggressive or malformed: random quotes, odd rephrasings, and occasional refusals with safety-style messages.
  • Grammar checks miss some issues (verb agreement, articles) that Grammarly and LanguageTool catch.
  • Others report it handles mixed-language sentences surprisingly well, and view it as a very promising first release.

Comparisons to Alternatives

  • LanguageTool and Harper are frequently mentioned; both have FOSS components and can be run locally (via Java, Docker, or Flatpak).
  • Several report LanguageTool with n-gram data is excellent; Harper is seen as weaker on basic errors but rapidly improving.
  • Some are building similar local grammar tools atop Chrome’s built-in LLM.

Security, Trust, and Licensing

  • Strong concerns about keylogging risk, especially for a tool that monitors all keyboard input.
  • Critics note: app is unsandboxed, distributed outside the Mac App Store, and appears to lack a clear corporate legal entity or independent audits.
  • Others respond that this risk exists for any proprietary app, but skeptics insist privacy claims need stronger technical and legal backing, or open source.

Performance, Platform & Integration

  • Uses an 8B Gemma 3n model (3 GB RAM); runs on Apple Silicon and offline. Some worry about RAM overhead.
  • Users report inconsistent operation in apps like Slack, VS Code, and browsers; system-wide, cross-app reliability is a key expectation.
  • Praised for being a local, one-time-purchase Mac app with a free demo, but many request Windows/Linux support and editor integrations (Vim/Emacs/API).

AI Detectors & Academic Use

  • One commenter worries LLM-style rewrites might trip AI-detection tools (e.g., Turnitin), making use risky for coursework.
  • Others argue that light grammar correction should be distinguishable from AI-generated text, but acknowledge academic policies often ban any LLM use outright.