Most checkers claim to be accurate. None says how it was measured. Korr publishes the corpora, the method and the numbers, including the ones that are not flattering.
Measured on 9 corpora of 40 sentences, in three languages. Half of every corpus is already correct.
| Corpus | Detection | Precision | Recall | F1 | Sentences damaged |
|---|---|---|---|---|---|
| en-courriel | 96.88 % | 100.00 % | 96.88 % | 98.41 % | 0 |
| en-technique | 88.89 % | 96.00 % | 88.89 % | 92.31 % | 0 |
| en-redactionnel | 100.00 % | 100.00 % | 100.00 % | 100.00 % | 0 |
| fr-courriel | 97.22 % | 97.14 % | 94.44 % | 95.77 % | 0 |
| fr-technique | 100.00 % | 97.56 % | 97.56 % | 97.56 % | 0 |
| fr-redactionnel | 94.44 % | 100.00 % | 94.44 % | 97.14 % | 0 |
| es-general | 95.74 % | 100.00 % | 95.74 % | 97.83 % | 0 |
| es-registros | 89.80 % | 100.00 % | 89.80 % | 94.62 % | 0 |
| es-tecnico | 95.74 % | 100.00 % | 95.74 % | 97.83 % | 0 |
98.67 % precision · 95.26 % recall · 96.91 % F1 · 0 sentences damaged
98.23 % precision · 95.48 % recall · 96.82 % F1 · 0 sentences damaged
100.00 % precision · 93.76 % recall · 96.76 % F1 · 0 sentences damaged
A checker that flags everything looks impressive and is useless: after the third wrong suggestion, you stop reading them. Korr only acts when it can prove the error. On 360 sentences, 174 of them already correct, it changed none of the correct ones.
Professional correspondence, product documentation, narrative and academic writing, in English, French and Spanish. Forty sentences each, written for this measurement.
Twenty sentences per corpus contain no error at all. They are the real test: any change to them is a false positive, and they are counted separately.
Each correction is aligned word by word against the expected text. A sentence with two errors and one fix counts as one hit and one miss, not as a failure.
The corpora ship with the project. One command replays the whole measurement, and the engine is deterministic: the same input always gives the same output.
These corpora were written by us. That is their limit and we state it: the sentences look like what the engine handles. The figure we stand behind is the precision and the zero damaged sentences, not the recall.
So we say nothing about it. Comparing would require running the same corpora through it and publishing that too. Until then, any claim would be unverifiable, and this page would be worth nothing.
Italian and Portuguese have their own test suites but no published benchmark yet. We would rather show three measured languages than five estimated ones.
| Korr | LanguageTool (free) | LanguageTool (self-hosted) | |
|---|---|---|---|
| Price | Free | Free tier + paid | Free |
| Character limit | None | 10,000 per check | None |
| Setup required | None open a page | None | Run a Java server |
| Works offline | Yes, by default | No | Yes |
| Text stays on your device | Always | No | Yes |
| Review each change | One by one | One by one | One by one |
| Personal dictionary | Yes, on your device | In your account | In your account |
| Account | Never | For premium | Not required |
| Paid tier holding features back | No | Yes | Premium rules excluded |
| System-wide shortcut (desktop) | Yes (Ctrl+Alt+C) | Browser + add-ons | Browser + add-ons |
| Languages | 5 FR, EN, ES, IT, PT | 30+ | 30+ |
Comparison as of September 2026. LanguageTool can be self-hosted, but that means installing and maintaining a server; its premium suggestions and n-gram data are not included.
On our published corpora, 98.67 % precision in English, 98.23 % in French and 100 % in Spanish, with zero correct sentence damaged out of 360. Precision means that when Korr changes something, it is right that often.
It is a correction applied to a sentence that was already right. It is the most expensive kind of mistake, because it damages text the writer had finished. Korr recorded none on the 174 correct sentences of the benchmark.
Because Korr abstains whenever it cannot prove the error. It would rather miss a mistake than invent one. Raising recall by guessing would lower precision, and that trade is refused by design.
Yes. The nine corpora are plain JSON files shipped with the project, and one command replays the whole run. The engine has no randomness, so the numbers do not move.
No. There is no analysis server. After the first load you can disconnect and everything still works, which is the simplest way to verify it yourself.
English, French and Spanish, three corpora each. Italian and Portuguese are supported and tested, but no benchmark has been published for them, so we claim nothing about their accuracy.
Yes. No account, no subscription, no quota, and no paid tier holding features back.
Yes, on the Chrome Web Store. It corrects text directly in Gmail, Slack or any text field, without copying and pasting, and it runs the same local engines: nothing is sent anywhere. It also installs in Edge, which accepts extensions from the Chrome Web Store.