Harper – an open-source alternative to Grammarly

Original Article ↗ Hacker News Discussion ↗

Architecture & Approach

  • Harper is a rule-based grammar checker written as a classic NLP system, not an LLM.
  • Uses a mix of hard-coded phrase corrections and more dynamic rules (e.g., your/you’re, Oxford comma, other syntactic checks).
  • Some see “decades of NLP research” in the marketing as overstated, given the current code looks like a relatively small, hand-built rule engine without obvious use of older statistical/NLP resources.

Performance, Integrations & UX

  • Key selling point: runs locally, in under ~10ms per document, and can be embedded into apps.
  • Ships with an LSP server; people like it for checking comments in code editors (Neovim, etc.), plus browser extensions including Firefox.
  • One report that the LSP consumed >1 GB RAM in Neovim.
  • Users request: keyboard shortcuts, delay while typing, rule import/export, iOS keyboard, Word and mobile Obsidian support, better Outlook web support, and a stable web demo.

Quality & Coverage

  • Mixed feedback: some find it a “good start” or “decent,” but many note missed basic errors (“Me and Jennifer went…”, “My name John…”, nonsense sentences) and occasional bad suggestions.
  • Several conclude it will need years of rule-writing to get close to Grammarly-level coverage.
  • English-only for now (US/UK/CA/AU variants); broader language support is “on the horizon.”

Comparison to Grammarly, LLMs & Other Tools

  • Many praise the non-LLM design: deterministic, fast, private, and less “AI-bloated” than Grammarly, which multiple users say has become inconsistent, buggy, and overly verbose.
  • Others argue a hand-written rule system is fundamentally inadequate in the age of LLMs; they want an open-source, LLM-based Grammarly clone, especially for multilingual use.
  • LanguageTool and Vale are discussed as existing open-source alternatives; some note Harper’s site should acknowledge that LanguageTool can also run locally.
  • There’s debate about grammar tools in general: some rely on them heavily (dyslexia, non‑native speakers, professionals needing proofreading), while others question why OS/browsers don’t just build this in.

Philosophy & Ownership

  • Long subthread on rule-based vs probabilistic models, language evolution speed, slang, and whether a stable rule set is desirable versus tracking real-time linguistic change.
  • Some distrust Automattic’s long-term stewardship but are reassured by Harper being FOSS and thus forkable.