HarkbackGetting started
Five minutes from nothing to your first remembered term.
1. Install
Harkback is a Chrome extension (Manifest V3). There is no store listing yet, so load it from source. You need Node 22 or newer and pnpm.
git clone https://github.com/zestworks-io/harkback.git
cd harkback
pnpm install
pnpm --filter @harkback/extension build
Open chrome://extensions, turn on Developer mode, choose Load unpacked and select apps/extension/.output/chrome-mv3. The onboarding page opens by itself.
After you pull new code, run the build again and press the reload icon on the extension's card.
2. Connect a model
Harkback has no server. It needs a model to write explanations, and you choose which one. The onboarding page is a three-step stepper; a finished step shows a ✓ and can be clicked to go back.
- Welcome. See an example reunion hint, and click the underlined term to see what an explanation looks like. No model is called.
- Privacy. Read where your text goes and tick the consent box. Next stays disabled until you do.
- Model. Pick a provider; the address fills in. Paste an API key if the provider needs one (Ollama and Chrome's built-in model do not). Press Test connection, and Chrome asks whether Harkback may reach that address; allow it. A spinner and "Connecting…" show while the test runs, and the button is off until it ends. Then press Finish.
Chrome built-in (Gemini Nano) needs no address or key. Selecting it shows the model's status and a Download model button; Finish waits until the download is done. It is a small model that runs on your computer, so expect it to be slower than a cloud model, especially on the first request, with simpler answers. It suits short explanations.
You can add more models and change options later on the settings page. Running onboarding again updates the model you already have; it does not add a second copy.
The easiest free option is a local Ollama model. It needs one extra step, because Ollama refuses requests from browser extensions until you allow the extension's origin. The Test button shows the exact command for your system. See Models and providers for every provider and for running a model on another computer in your home network.
The interface starts in English. You can change its language, and the language of the explanations, on the onboarding or settings page.
3. Explain a term
- Open a paper on arxiv.org. arXiv pages are scanned automatically.
- Select a term, for example
LoRA. - Click Explain next to the selection, or press
Alt+Shift+E. - Read the answer as it streams in. A green tag says the page itself defines the term; a yellow one says the answer comes from the model's own knowledge.
- Press Got it or Still confused. That answer is the first grade for the term and decides when Harkback brings it back for review.
Other ways to start: the right-click menu on selected text, and a selection made with the keyboard or by touch.
4. Meet the term again
Open a different paper that mentions LoRA, low-rank adaptation or LoRAs. The term is underlined. Hover over it to see when and where you looked it up, and what you understood then. You can mark it remembered, explain it again, compare the two usages, or mute it.
5. Look at what you have
Open the History page (linked from settings and onboarding). It lists every term by concept, with search, follow-up conversations, a review queue and a graph of how your concepts connect. From there you can also export your records and restore them from a backup.
Reading something that is not on arXiv
- A web page: click the toolbar button once. To scan a site automatically from then on, add it under Sites in settings.
- A PDF: click the toolbar button on the PDF. arXiv PDFs open as their HTML version; other PDFs open in Harkback's own reader. A PDF on your computer needs two approvals the first time; see Troubleshooting.
- Something confidential: mark the site or the source as sensitive first. See Privacy and sensitive sources.