Choosing your AI
Your companion’s mind is a language model that you select. Unclaw ships no locked-in AI and takes no cut of your usage. You connect a provider directly, and your conversations flow between your machine and that provider only.
There are three routes, and Check Keys verifies whichever you pick before you’re done.
Route 1: an API key
Section titled “Route 1: an API key”Paste a key from any supported provider and choose a model:
| Provider | Notes |
|---|---|
| Groq | Extremely fast, generous free tier: the easiest great default |
| OpenAI | GPT-5.x family |
| Anthropic | Claude models |
| Google Gemini | Gemini models; also powers grounded web answers |
| DeepSeek | Strong value |
| Cerebras | Very fast inference |
| Others | The provider list in Settings is current |
Keys live in your system’s secure credential storage and go only to their own provider (details).
Route 2: a subscription you already pay for
Section titled “Route 2: a subscription you already pay for”If you subscribe to Claude, ChatGPT, or Gemini through their coding CLIs (Claude Code, Codex, Gemini CLI), Unclaw can use that subscription as the mind: no API key, no extra bill. Sign in to the CLI once on your machine; Unclaw detects it and lists its models.
Route 3: fully local with Ollama
Section titled “Route 3: fully local with Ollama”Run a model on your own hardware with Ollama and your companion’s mind never leaves the room. Smaller local models trade some depth for total privacy and zero cost, and they pair naturally with the local voice engines for a configuration where conversation never touches the internet at all.
One model does everything. With Tools on, heavy asks (real research, web tasks, anything that needs a browser, files or code) are handled by the same model you picked above: it takes its time, drives a browser, searches the web, reads Wikipedia, works with files and keeps a working memory, narrating its progress out loud while it works. Quick replies stay quick. Sharing images with your companion also goes through this path when the model can see.
- Works with any cloud provider, the CLI subscriptions, and local Ollama models that can call tools (marked Tools in the picker).
- Thinking on tasks sets how hard the model reasons while it works, separately from the quick conversational turns.
- Off by default: with the toggle off, she simply answers conversationally and never reaches for a tool.
Switching and mixing
Section titled “Switching and mixing”Change provider or model any time in Settings (mid-relationship, even); her memory of you carries across minds. A mid-size model such as Claude Haiku or Sonnet handles both the quick turns and the tool work well.