Models & providers
Start free with no API key, add your own keys when you need more, or go fully local with Sonderr-v1. Your workspace, your choice.
Kilo Gateway — free AI models, no API key
Kilo Gateway is the default provider on a new install. Choose it in Settings → API & Models, discover the current free-tagged models, and select one to start without entering an API key. Anonymous requests are rate-limited, and Kilo's free-model lineup can change — adding a Kilo API key unlocks the account's broader catalog.
Sonderr still owns the system prompt, skills, tool selection, permission checks and local tool execution — the gateway supplies model responses. Prompts and context are sent to Kilo over HTTPS; this is hosted AI, not local inference. See Kilo's current authentication and free-access rules in their official docs.
Sonderr-v1 — the local model
Sonderr-v1 is Sonderr's first small language model: 0.6B parameters, specialized for the Sonderr environment. Choose it at the top of the model picker. On compatible Linux x64 systems (glibc 2.34+, GLIBCXX 3.4.31+, OpenSSL 3 and libgomp), the installer asks before downloading the SHA-256-verified Q4_0 GGUF model and bundled llama.cpp CPU runtime — about 370 MB total. Other platforms use the merged model with the CPU-only PyTorch/Transformers runtime.
Model weights and runtime files stay under ~/.sonderr; inference runs locally and nothing leaves your machine. The cloud-hosted option is not available yet. While Kilo is selected as the provider, picking Sonderr-v1 from the model picker keeps Kilo chosen while its inference requests run through Sonderr's own local runtime.
Choosing models
- Settings → API & Models is the control room: providers, keys and the model catalog.
- Model discovery is provider-agnostic, with ranking and vision-capability flags per model.
- The picker filters to image-input models when you switch into Vision mode, auto-switching if the active model cannot see.
What goes where — the honest map
| Path | What is shared | What is not |
|---|---|---|
| Kilo Gateway (hosted models) | Prompts and context, over HTTPS | Your files stay local; tools execute on your machine |
| Sonderr-v1 (local model) | Nothing — inference is local | No network calls for inference at all |
| Release check (GitHub) | A normal HTTPS request to GitHub | Sonderr does not send your version, prompts, files, settings or credentials |
| Web research tools | Bounded read-only public HTTPS requests | Local/private hosts are blocked; pages are size-bounded |
The version comparison for required updates happens on your machine — GitHub can see the request's source IP and metadata, and nothing else about your workspace travels with it.