AI Grammar Checker
Fix grammar and spelling errors in pasted text. Runs locally via transformers.js — your draft never leaves your browser.
Why a Grammar Checker That Runs Locally
Hosted grammar tools see every sentence you check. For confidential drafts, legal documents, internal email, journalism work, or anything else you would not paste into a public web form, that visibility is a problem. This tool runs a T5-based grammar correction model — pszemraj/grammar-synthesis-small, about 80 MB on disk — entirely in your browser. The text you paste stays on your machine. The model is fine-tuned for sentence-level grammar correction: typos, missing articles, subject-verb agreement, incorrect tense, doubled words. It handles light-to-moderate errors well. Heavier rewrites — restructuring a paragraph, switching from passive to active voice across multiple sentences — are out of scope and better suited to the Paraphraser tool. Released under Apache 2.0 by pszemraj on Hugging Face, the model is free for commercial use. Trained on a synthesized augmentation of the JFLEG dataset, which is the standard benchmark for grammar correction.
How the Grammar Checker Works
First-visit downloads the grammar-synthesis-small model — about 80 MB, cached in your browser. Paste your text in the input area. The tool splits the input into sentences using a punctuation-aware splitter (so abbreviations like "Dr." and "U.S." do not split mid-sentence), then feeds each sentence to the model. Corrections appear inline in the output panel with diff highlighting: words removed are struck through, words added are underlined. A copy button copies the corrected text without the diff markup. The model is conservative by design — it prefers leaving text alone over changing it incorrectly, which means it sometimes misses real errors but rarely introduces new ones. For aggressive rewriting use the Paraphraser. The model is English-only and trained primarily on American English; British spellings may be flagged as errors. Sentence length over 128 tokens (about 100 words) gets truncated, so for long sentences split at commas or semicolons before pasting.
Frequently Asked Questions
Which grammar correction model powers this tool?+
The Xenova ONNX port of pszemraj/grammar-synthesis-small — a T5-based model fine-tuned on a synthetically-augmented version of the JFLEG benchmark dataset. Approximately 80 MB compressed, released under the Apache 2.0 license.
Is my draft sent to a server for grammar analysis?+
No. After the initial model download, every correction runs entirely in your browser. The text you check is never sent to a server, including ours. No request logs, no retention policy.
What kinds of errors does it catch?+
Typos, subject-verb agreement, missing or misused articles, incorrect tense, doubled words, common preposition mistakes, and basic punctuation. It does well on sentence-level errors. Paragraph-level restructuring or style rewriting is out of scope — use the Paraphraser tool for that.
What kinds of errors does it miss?+
Stylistic choices (long sentences, weak verbs, jargon), structural issues across multiple sentences, factual or logical errors, and very subtle agreement errors. The model is conservative by design — it would rather leave a sentence alone than correct it the wrong way.
Does it work for British English, or only American?+
Primarily American English. British spellings like 'colour' and 'organise' may be flagged as errors. We do not currently have a clean-licensed British English grammar model on the Xenova side; if one ships we will add it as a toggle.
Can I check very long documents?+
Yes, but the tool processes sentence-by-sentence, so a 5,000-word document takes proportionally longer than a 500-word one. For very long inputs consider running them through the Readability Score tool first to find the rougher sections, then grammar-check only those.
Why does the same sentence sometimes get a different correction?+
The model uses greedy decoding by default, but small variations in surrounding context (a comma here, a period there) can shift its output. For consistent results across runs, paste the entire passage at once rather than splitting it across multiple sessions.
How does this stack up against Grammarly or LanguageTool?+
Grammarly is much more sophisticated — it covers style, tone, clarity, and inclusive language alongside grammar, and is fine-tuned on much more data. LanguageTool is rule-based and catches a different mix of errors. This tool is narrower (sentence-level grammar only) but ships with zero upload, no signup, no premium tier, and works fully offline after the first model download.
Built by Derek Giordano · Part of Ultimate Design Tools
Privacy Policy · Terms of Service