Named after the original AI hunter in Blade Runner, Deckard is a Chrome extension that detects AI-generated text on pages you visit.
Install Deckard v0.6.4 from your terminal:
curl --proto '=https' --tlsv1.2 -fsSL https://github.com/sgoedecke/deckard/releases/download/v0.6.4/install.sh | bashIn Chrome:
- Open
chrome://extensionsand turn on Developer mode. - Choose Load unpacked and select
~/Deckard/extension(the Deckard folder in your home folder). - Pin Deckard so you can see it in your extension hotbar
Right now this only works on Apple Silicon macs. If you want to use it on a PC or some other device, PRs are welcome.
The installer defaults to ~/Deckard; --home DIR selects a different location.
Older installers use ~/Library/Application Support/Deckard. To move an existing
installation, first run
"$HOME/Library/Application Support/Deckard/current/bin/deckard" uninstall,
remove the old extension in Chrome, then install with the new installer and load
~/Deckard/extension. Do not move the folder manually: the native-host registration
and shell PATH refer to its installed location.
Deckard downloads and runs the Gradient model on your laptop. This will consume a few hundred MB of memory while you're browsing. When you visit a page, the extension will chunk it and run it through the local model.
This is obviously much less reliable than Pangram (which at the time of writing is the only good AI detector), but (a) it runs entirely locally, and (b) you can use it as much as you want for free.
I've set the default detection threshold to just above 98%, which gives a low false-positive rate, but that's configurable via the extension slider. Please don't use this as proof of AI usage; if you want to do that, paste it in to Pangram.
Deckard only scans in 50 word chunks, so short AI-generated content is harder to detect (since it'll be chunked with other text on the page).
Run deckard uninstall (or
"$HOME/Deckard/current/bin/deckard" uninstall
if PATH was not configured). It removes Deckard's managed installation files,
owned PATH block and native-host registration. Then manually choose Remove
for Deckard at chrome://extensions; the CLI cannot remove a Chrome extension.
This (aside from the README above this point) is entirely vibe-coded, so contribute by hand at your own risk.
Use a current Node.js with the built-in test runner:
npm testExtension tests need no npm dependencies. Native tests/builds require the native toolchain and fixtures; release users do not. On a supported Mac, with a prepared dependency cache and canonical quantized model directory:
NATIVE_CACHE=/path/to/native-build scripts/build-native.sh
DECKARD_MODEL_DIR=/path/to/canonical/mlx-q4 npm run test:native
scripts/package-release.sh --model-dir /path/to/canonical/mlx-q4The build reads the external dependency cache without modifying it. Packaging
also accepts --native-dist native-cli/build/dist and
--output-dir dist/v0.6.4. These are maintainer steps, not evidence that a
release has been published. Release archives bundle prepared model assets;
source installs must pass --model-dir explicitly. The installer does not
download or convert upstream FP32 weights automatically.
Without DECKARD_MODEL_DIR, native model-installation tests are explicitly
skipped; protocol and ownership-refusal tests still run against the built CLI.
The separately invoked native-cli/tests/model-smoke.mjs uses original synthetic
text, serial inference, a 6 GiB physical-footprint watchdog and a 90-second
deadline. Pass an installed CLI path and a new receipt path; the developer build
provides its process meter, or set DECKARD_PROCESS_METRICS explicitly.
Load extension/ unpacked for extension development;
its manifest key gives the same ID, so do not load both copies in one profile.
See native protocol for framing and model identity. Research datasets, experiment outputs and model weights are not committed.
Deckard is MIT licensed. Gradient and its Microsoft DeBERTa-v3-large base are MIT-licensed upstream.