Connecting an AI provider
Presets, keys, the connection test, and how to read the errors.
NovelKnow isn't tied to any model vendor — you bring your own API key, and which model you use and what it costs is yours to decide.
The fast path
AI settings → click a provider preset → paste your key → test the connection → save.
The preset fills in the base URL, default model and context window. Those are the three things people most often get wrong, and the usual reason a setup doesn't work.
Built-in presets: DeepSeek, OpenRouter, OpenAI, Kimi / Moonshot, Qwen / DashScope, SiliconFlow, and local Ollama.
Any OpenAI-compatible endpoint works — pick the custom option and type the base URL yourself.
Model lists
After saving, "同步模型" pulls the real model list from the provider. From then on the model dropdowns are populated from it, so you never type a model name by hand.
Mark the ones you use often as favourites and they sort to the top.
Reading errors
Connection tests and generation failures include the provider's own response, so you can tell these apart:
| Error contains | Usually means |
|---|---|
401 / Incorrect API key | Wrong or expired key |
404 / model does not exist | Model name is wrong, or your account lacks access to it |
429 | Rate limited, or the account is out of credit |
fetch failed / ECONNREFUSED | The server can't reach the provider — see the proxy note below |
Outbound proxy
If your deployment can't reach a provider, enable the outbound proxy in the AI settings dialog.
That address is resolved from the deployment server, not from your own computer. Pointing it at 127.0.0.1:7890 only means something if a proxy is running on the server itself.
How your key is stored
Keys are encrypted with a key derived from your account and only returned to the browser masked. You will never see your own key in full again — to change it, paste a new one.