Margin: AI second pass for Harper

Catches what Harper's rules can't, in web text fields. Local model (Ollama, LM Studio) or Gemini. Not affiliated with Harper.

Margin is an optional AI companion to Harper, the private grammar checker. Harper's fast rules stay in charge of spelling and mechanical grammar; Margin adds what rules can't catch: agreement across clauses, tense, wrong-but-real words ("effect" for "affect"), missing or repeated words, run-ons and comma splices. It never repeats what Harper already flags. Use it together with Harper's own browser extension. - Purple underlines in text fields; click one to Apply or Ignore. - Rewrite a selection (Alt+Shift+R or right-click): Improve, Concise, Formal, Friendly, Fix only, or your own instruction, with Accept / Reject. - Off on every site until you turn it on for that site. Where your text goes: with a local model (Ollama or LM Studio), nothing leaves your computer. With Google's Gemini API and your own key, the text being checked is sent to Google; on the free tier Google may use it to improve its products. Code, URLs and emails are masked first. No account, no server, no telemetry. What's tested: Gemma 4 26B A4B, both on your own computer with Ollama and through the Gemini API. On 747 learner sentences corrected by people, Harper alone catches 24% of the fixes and Harper + Margin 40%; Margin changed none of the sentences the people left alone. In this extension, on 20 real GitHub issues, suggestions appeared 4 to 7 seconds after typing. The recommended local model needs about a 16 GB GPU (ollama pull gemma4:26b-a4b-it-qat); the options page shows the one-time setting Ollama needs for browser extensions. LM Studio and other models aren't tested yet; reports welcome on GitHub. Not affiliated with Harper or Automattic. Open source (MIT): https://github.com/phoenixak/margin