Automattic's grammar checker skips the API call entirely

Harper runs a full grammar check without calling a model. Automattic's checker is written in Rust, runs entirely on the user's machine, and has picked up 15,183 stars, including one day that added 147.

That design answers a specific worry. Anyone who has hesitated to paste a draft into a chatbot, because it names a client or quotes an unreleased price, gets a checker that never sends the text anywhere. No API key, no per-request cost, no dependency on a model staying online.

The repo's own numbers show what that independence costs. Harper has 890 open issues against 614 forks, and that ratio is the tell. A model-backed grammar checker inherits whatever the underlying model already knows about English. Harper doesn't get that for free. Every rule, every exception, every case of English actually getting written wrong has to become someone's pull request. The open issues are the backlog of sentences the checker still misses.

That's the actual shape of the trade-off between local and API-backed tools, laid out in a public issue tracker instead of a pitch deck. A model-backed checker ships broad coverage on day one and drifts when the model underneath it changes. A hand-coded checker like Harper starts narrower and improves only as fast as its maintainers and contributors close that backlog.

Picking a grammar tool for anything sensitive comes down to that trade. A vendor's model reads the draft in exchange for broader coverage. Harper's contributors read the issue queue instead.