Reads the pull request before you do, and tells you how much of it sounds like a language model wrote it.
Install: addons.mozilla.org. Free, no account, 30 scored comments a day.
Click scan on any pull request or issue. Every comment gets a badge, and every badge opens to show the actual signals behind it: which buzzwords fired, which stock phrases, how uniform the sentence lengths are. On a repo's pull request list you can scan up to 25 open PRs at once.
Codeberg, Godot, OpenJDK, Rust and Ghostty all restricted AI-generated contributions during 2026. Whichever way a project lands on it, enforcing the decision still means a human reads every submission and forms a private opinion about who wrote it. This does the first read and shows its work.
It is a triage tool. Nobody can prove authorship of text after the fact, including this, and a maintainer who treats a red badge as a verdict will eventually accuse a real person of something they did not do.
Detectors get sold on a single accuracy number that falls apart the moment the text is not from the generator they tuned on. Here is the whole picture instead, measured on held-out sets:
| Held-out set | What it is | AUROC | Catch rate at 5% FPR |
|---|---|---|---|
| DetectRL-X | 2025-era generators (DeepSeek-V3 and similar), plus paraphrase and zero-width-space attacks | 0.983 | 92% |
| RAID-test | 2022-23 generators (GPT-2/3, Cohere, MPT, Mistral) | 0.89 | 67% |
| MAGE (paraphrase) | GPT text run through a paraphraser, out of domain | 0.81 | 39% |
| humanizer-attack | Ghostbuster "undetectable" essays | 99% caught |
The number that matters most is the one that goes the other way. On 4,300 held-out essays by non-native English speakers, real writing by real people, 7.8% come back flagged. That is the honest ceiling on trusting any single result, and it is why the advice never changes: judge the contribution, not the author.
The threshold targets a 5% false-positive rate across a broad mix of English. Under roughly twenty words, or on text it reads as anything other than English, it does not guess at all. It says it abstained and tells you why. A one-line "lgtm" gets no verdict, because no honest detector can give one. The measurement page carries the corpus, the calibration and the named failure modes.
Nothing, until you click scan. Then the visible text of that PR or issue goes over HTTPS to the scoring worker with code blocks stripped, plus a random id that counts your daily free scans and rotates every day. The text is scored in memory and dropped. It is not stored, not logged in full, and not tied to your GitHub account.
The extension asks for one permission, storage, and runs on github.com only. If you add
a GitHub token to raise the batch-scan rate limit, it stays on your device and is sent only
to github.com/ghapi. The privacy policy is the
long version, and the code that backs it is in this repo.
extension/ the MV3 extension, Firefox first
content/ the scan button, badges and signal panels on github.com
background.js every network call lives here
popup/ options/ quota, CSV export, settings
worker/ the scoring API on Cloudflare Workers
src/index.js POST /score, metered per install and per IP in a Durable Object
site/ the landing page, measurement page and privacy policy
test/ worker suite and a headless DOM smoke test
python3 tools/sync_detector.py # pull the private detector into the worker
node --test test/worker.test.mjs # 23 tests, no network
python3 test/content_smoke.py # 16 checks, drives the real content script in headless chromium
bash tools/package.sh # lint and build dist/
To try a build: Firefox, about:debugging, This Firefox, Load Temporary Add-on, pick
extension/manifest.json.
Deploying the worker needs a Cloudflare account:
cd worker
wrangler kv namespace create QUOTA # paste the printed id into wrangler.toml
wrangler deploy
Then put the live URL in extension/config.js if it differs from the default.
The detector. worker/src/detector-core.js is synced in at build time from a private
engine and is gitignored, so the worker suite does not run in CI here either. Everything
about how that engine is called, metered, rate limited and rendered is in this repo, which
is the point of the privacy policy telling you to read the code. The claims about what the
server does with your text are checkable. The model is not open.
vendor/ExtPay.js is the unmodified ExtensionPay client, staged for a paid tier that does
not exist yet. It is not in the manifest and not in the shipped zip, so the extension you
install carries no payment code, no extensionpay.com permission and no upgrade button.
Every score says signal, not proof. Do not weaken that framing anywhere in the UI. A false accusation costs a contributor far more than a missed bot costs a maintainer.
MIT, see LICENSE.