What BYOK actually means
ScreenNote's AI works in two ways. One is the official AI: you claim it inside the app, it comes with a monthly free allowance, and after that a Pro subscription continues it. The other is the BYOK route this piece covers: you go to the website of whichever AI provider you choose and request an API key (essentially your own personal key), enter it into ScreenNote's settings, and from then on every AI call is billed directly between you and that provider — pricing, quotas, and privacy policy are all that provider's own rules, and ScreenNote neither handles the money nor adds a markup.
The upside is transparent pricing and no lock-in to a single vendor — you can switch anytime. The cost is that you have to go request a key yourself first, one extra step beyond "just open it and go," though that step usually takes only a few minutes.
17 supported providers, plus 1 universal slot
The full list: Kimi (Moonshot AI), OpenAI, DeepSeek, Zhipu (GLM), Qwen (Alibaba), Anthropic, Stepfun (阶跃星辰), Doubao (ByteDance), Tencent Hunyuan, SiliconFlow, 01.AI (Yi, 零一万物), Baichuan, Google Gemini, Mistral, Groq, xAI, and OpenRouter.
There's also a "custom OpenAI-compatible" universal slot on top of that — any service that follows the OpenAI API format can be connected by manually entering its endpoint, even if it's not on the 17-provider list above. This slot doesn't count toward the 17; it's there as a fallback for "I know of a service you don't directly support."
Before you choose, know exactly what the AI is doing here
By default, what ScreenNote sends to AI is the text local OCR has already recognized, so a plain text chat model is entirely sufficient and there is no need to pay extra for multimodal capability; just pick that provider's cheapest text-only tier. If the model you choose supports image understanding, the app may also send the screenshot image to read it more accurately. Whether that happens depends on the model you pick, and the app asks for your consent when you first save a key. See section 4 of the privacy policy.
4 practical factors to weigh
- Network reachability: if you're inside mainland China, overseas providers like OpenAI, Anthropic, Google Gemini, Mistral, Groq, and xAI may need extra network conditions to connect reliably. DeepSeek, Zhipu, Qwen, Kimi, Doubao, Tencent Hunyuan, SiliconFlow, 01.AI, and Baichuan all connect directly from within China — configure one and it just works, no need to troubleshoot connectivity first.
- Chinese-language quality: most of what ScreenNote processes are Chinese-language screenshots (chat logs, bills, articles). Domestic models tend to understand Chinese context and write summaries in a style that fits more naturally — the titles and summaries they generate read more like something a person would actually write, less stiff.
- Cost per call: this is a high-frequency, low-volume scenario — one call per screenshot, and the text is usually short. There's no need to reach for the most expensive, most powerful flagship model; every provider's entry-level or lightweight tier is typically cheap enough that you won't feel any cost pressure running it. Aggregator platforms like SiliconFlow and OpenRouter also let you switch between multiple models under a single key, making it easy to compare.
- Whether you're already using one: if you've already activated a given provider's API for something else — say, a script you wrote that calls DeepSeek or OpenAI — just reuse the same key. No need to open a brand-new account just for this one app.
Setup takes just 3 steps
- Go to your chosen provider's console and request an API key (most have a low barrier with no phone-number ID verification required, and it takes just a few minutes);
- Open ScreenNote's Settings → AI Settings, select the matching provider, and paste in the key;
- Use the "Test Connection" feature to confirm the call works — from then on, every newly imported screenshot automatically runs through AI processing.
For exactly what capabilities configuring AI unlocks — like one-tap copying of verification codes and order numbers — and what's still left on the local side if you don't configure it, see this piece on using it without AI.
