Tutorial · Models and usage
Models and Usage
Foreverse is BYOK-first at the model layer: you bring your API key and connect directly to the provider, with keys AES-encrypted and stored only on this device. If you don't configure a key, you can still use the official channel with pay-as-you-go billing. This guide covers the full setup flow and how to read the "API request logs".
Add providers and keys
- Entry: Settings -> AI Models & Services. 60+ preset providers (OpenAI / Anthropic / Gemini / DeepSeek / xAI / Alibaba / MiniMax...) include real logos and search.
- Open a provider: paste the key (masked when displayed) -> Test connection. The app sends one real request, shows round-trip latency, and fetches the model list automatically.
- Custom provider: any OpenAI / Anthropic-compatible endpoint or self-hosted proxy can be added. Choose OpenAI-compatible or Anthropic protocol; the API URL must include the full version prefix, such as
https://api.example.com/v1. The app won't guess it for you. - Security boundary: keys are encrypted, stored only on device, never uploaded, and never written to logs; requests go directly from your phone to the address you configured.
Enable models and set defaults
- Each provider page lists available models, which you can enable or disable one by one. Only enabled models appear in model pickers. You can also add a model ID manually.
- Default model configuration (by modality): set defaults separately for text / image / speech / video. Continue writing uses the text default, image generation uses the image default, listening uses the speech default, and they don't interfere with each other.
- Every AI entry point (continue form, chat, Agent) can switch models temporarily without changing global defaults; each entry remembers its last choice.
Official channel
Sign in to use built-in models with pay-as-you-go billing: 1 credit = $0.0001, deducted from prepaid balance, with no subscription. Official and BYOK can be mixed. For example, text can use your DeepSeek key while image generation uses the official channel. You can switch any time; your data stays put.
API request logs
Settings -> API request logs. Every AI call is recorded:
- Top stats: request count, success rate, total input/output tokens, and cache-hit footnote (hit count · cache tokens · hit rate).
- Filters: BYOK / official channel and modality (text / image / speech / video).
- Rows: provider logo + model name + input/output/cache tokens + time + success/failure status dot, grouped by day.
- Details: full token split (including cached input / cache miss / cache write), provider cache hit and discount labels, latency, and official-channel credit charges.
If you continue writing twice in a row, the second call's "cached input" should be close to total input. If it stays 0, something is breaking prefix caching. The cost-saving principle is explained in this test write-up.