Automattic/harper

The problem

Anyone who writes for a living has to decide between sending their drafts to a cloud grammar service or proofreading by eye. Cloud checkers add a network round-trip to every revision and mean every paragraph leaves your machine. Harper is an offline alternative: a Rust grammar engine that lints locally in milliseconds and ships as a language server, an Obsidian plugin, editor extensions, and a WebAssembly build for the browser.

What's different

The obvious comparison is Grammarly, which is closed, cloud-hosted, and priced as a subscription. LanguageTool is the open alternative, but it is memory-hungry and needs a large n-gram dataset to be useful, which makes it slow on ordinary documents. Harper's bet is the opposite: keep the rule set small and the engine fast, accept English-only support, and embed it where the text already lives. The trade-off is that a rule-based checker without a huge statistical model will miss things a cloud model catches; Harper argues that fewer, more contextual suggestions are better than many wrong ones. Its other distinguishing trait is packaging — the same core runs in editors via harper-ls, in Obsidian, and in the browser via WebAssembly, so you do not adopt a new writing app to get the checker.

Who it's for

Writers, students, and non-native English speakers who care about privacy and latency, and developers who would rather get suggestions inside Neovim, Helix, Zed, or Emacs than in yet another web editor. It is also a reasonable fit for teams that cannot send text to third-party servers. Look elsewhere if you need multi-language support, or if you expect the breadth of a large language model and want it to rewrite sentences rather than flag grammar.

Cost of adoption

Apache-2.0, so commercial and embedded use is fine. Installation depends on the integration: the README lists docs for VS Code, Neovim, Helix, Emacs, Zed, Obsidian, and harper.js, and points to Crates.io and npm instead of giving a single command. The main costs are scope and maturity — English only, and the README is thin on benchmark methodology and on how rules are maintained, so expect to judge the suggestion quality yourself. There is also a maintenance question: it is backed by Automattic but ships many separate integrations, each of which can lag.

Verdict

Harper is a real project rather than a weekend spike: over 1,000 days old, Apache-2.0, 17k stars, and a recent weekly Rust board appearance at #9 with roughly 500 stars in a week. The Hacker News footprint is tiny (a 4-point thread), so most of that attention is not coming from sustained public discussion. That pattern — a long-lived repo with a sharp weekly spike — suggests an established codebase that just got discovered, not a new project proving itself. If offline, low-latency proofreading fits your workflow, it is worth trying for an afternoon; if you need multilingual or model-grade rewriting, keep waiting.