Models and providers

Harkback sends one request to the model you configure whenever you ask for an explanation or a follow-up. It has no server of its own, and nothing else is sent anywhere.

Providers

Picking a provider fills in its address; you can still edit it. The provider decides the wire format.

An API key works only in the region where it was created: if a key from the other region of Qwen, Moonshot or Zhipu is refused, edit the address to that region's (dashscope-intl.aliyuncs.com, api.moonshot.cn, api.z.ai); Harkback still recognises it as the same provider. Choose Custom for LiteLLM, vLLM, LM Studio, llama.cpp's server, a company gateway, or any other service that speaks the OpenAI chat-completions format. You may paste an address that already ends in /chat/completions or /messages; Harkback trims it.

The first time you test or save a model, Chrome asks whether Harkback may reach its address. If you decline, the explain card says so instead of failing with a network error.

Which address is allowed

"Local" matters because sensitive sources are only ever sent to a local model. A server on your home network is reachable only from inside it, so plain http is accepted, but it is a different machine, so it is treated as remote.

Ollama

Ollama rejects requests from browser extensions unless the extension's origin is allowed. The Test connection button shows the exact command with your extension id. For example, on macOS:

launchctl setenv OLLAMA_ORIGINS "chrome-extension://<your-extension-id>"

Then quit and restart Ollama. On Linux with systemd, add Environment="OLLAMA_ORIGINS=chrome-extension://<id>" to the service with systemctl edit ollama. On Windows, set OLLAMA_ORIGINS as a user environment variable and restart Ollama.

To use Ollama on another computer, start it with OLLAMA_HOST=0.0.0.0 there and enter http://<that-computer>:11434/v1 here. Ollama has no authentication, so only do this on a network you trust.

Small local models may write weaker concept cards (relations and aliases). That lowers recall quality but never stops an explanation from being recorded.

Chrome's built-in model

Chrome built-in (Gemini Nano) runs on your computer through Chrome's Prompt API. It needs no address or API key, and counts as a local model, so a sensitive source can use it. Chrome only starts the download from a click, so use Download model in onboarding or settings; until it is ready a request says so instead of failing.

It is a small model. Expect it to be slower than a cloud model, especially on the first request, and its answers to be simpler. It suits short explanations; use a cloud or Ollama model for longer ones. It works best in English, Spanish and Japanese (German and French are also accepted), and its small window means a prompt that does not fit is reported, not cut.

Several models

You can configure as many models as you like.

• Default model – used for everything unless a rule says otherwise.
• Sensitive provider – tick this on one local model. Sensitive sources use it, and never any other.
• Per site – a site rule can name a model for matching pages.
• After a failure – the explain card offers each of the other models as Try with ….

Rate limits

Harkback limits itself to 10 explanations per minute and 100 per hour by default (settings → Rate limits). A request that fails before an answer arrives does not count.

API keys

Keys are stored unencrypted in the extension's storage. Other sites and extensions cannot read them, but anyone with access to your computer's disk can. A key is only sent to the address you entered, and never goes into backups or logs. Use a key with a spending limit, or a local model.