Paste your text and press Check grammar. The corrected version appears above a list of the individual changes, and nothing you paste leaves this device - Chrome with built-in AI checks with zero download, and every other modern browser can load a small open model once and keep it cached. A sample with deliberate mistakes and its corrected version are already on the page, so you can see the output shape before touching anything.
AI Grammar Checker - Private, In Your Browser
Fix grammar, spelling, and punctuation without sending your text to anyone. The check runs on your own device: Chrome versions with built-in AI correct with zero download, and every other modern browser can load a small open model with one click. Nothing you paste is uploaded anywhere, on any tier.
What you get back
The result is the corrected text plus a change list: every fix is shown as the original wording, the replacement, and a short reason, so you can accept each correction knowing exactly what moved. Your wording is preserved - this page fixes errors, it does not rewrite your tone or style. One click on Copy result puts the corrected version on the clipboard.
Three ways the check can run
The page always picks the cheapest honest path and says which one it used.
| Engine | Download | When it applies |
|---|---|---|
| Built-in browser AI (Proofreader) | 0 MB | Chrome versions that ship it - the page detects it and never assumes it is there |
| On-device open model | Size shown before download (Llama 3.2 1B default; 3B and Phi 3.5 mini as upgrades) | Any modern browser with WebGPU; weights are cached after the one-time download |
| Prepared sample demo | 0 MB | Every browser, instantly on page load |
The status line under the button names which engine produced the output currently on screen - built-in browser AI, the named on-device model, or the prepared sample - so you always know what you are reading. Model weights load only after you press Load model, with a progress bar and the download size stated next to each model before anything starts.
Honest limits before you rely on it
The on-device models are small (1-3 billion parameters). They catch common errors - agreement, tense, its versus it's, capitalization, everyday punctuation - but they stay below professional cloud checkers, so treat the output as a strong draft pass, not a final edit. On the model path, input is capped at 8,000 characters and the cap is stated right in the tool. Corrections are validated for English text, and there is no document upload - paste the text itself.
Where your data and the downloads live
Text stays in the page while you work and is not stored beyond the session; there is no account and no signup. The model cache and your last model choice are shared with the Private AI Chat page, so loading a model once serves both pages, and a Clear cached models button removes the downloaded weights whenever you want the disk space back.
Related tools on this site
For open-ended writing help in conversation form, the Private AI Chat runs the same on-device models without a fixed task. To shorten a text rather than correct it, the Text Summarizer condenses articles and notes with the same private, in-browser approach.
Frequently Asked Questions
Is my text really never uploaded?
Yes, on every tier. The check runs either through the browser's own built-in AI or through a model that was downloaded to your device - there is no server on this page for your text to be sent to.
Do I have to download anything just to try it?
No. A sample with deliberate mistakes and its prepared correction are visible the moment the page loads, and on Chrome versions with the built-in Proofreader the check on your own text also needs no download at all.
How big is the model download, and how often do I pay it?
The exact size appears next to each model in the picker before the download starts - the default Llama 3.2 1B is the smallest of the three options. It is a one-time cost per machine: the browser keeps the weights cached, and the same cache serves the Private AI Chat page too.
How good are the corrections compared with Grammarly-class tools?
Lower - and the page says so. Small 1-3B on-device models reliably catch everyday errors like agreement, tense, and its versus it's, but they miss subtleties a professional cloud checker finds. The change list exists precisely so you can review every fix before accepting it.
What happens with a very long text?
On the on-device model path the input is cut at 8,000 characters so it fits the model's context window; the tool states this cap and tells you when a text was truncated. Check longer documents in sections.
Can it also reword or improve my style?
No - it deliberately keeps your wording and only fixes errors. If you want open-ended rewriting or writing suggestions, use the Private AI Chat page, which runs the same on-device models without a fixed task.