AdroitPadOne interface, several backends
An assistant becomes difficult to maintain when every provider leaks into the user interface. AdroitPad places providers behind a common model interface: send messages, stream text, expose tool capability, and report errors using the same shape even when the underlying APIs differ.
That abstraction can cover Ollama, Claude, OpenAI, Grok, and custom OpenAI-compatible endpoints. Provider-specific details remain in adapters, where authentication, model identifiers, context limits, and streaming quirks belong.
Ollama as the default
Ollama is the privacy-first starting point because it needs no provider API key and can run without a network connection. Switching to Claude or OpenAI is an intentional choice from the provider control, not a hidden fallback that silently sends the current conversation elsewhere.
The UX should remain stable when the backend changes. Conversation history, tool status, cancellation, and errors need one presentation; only the capability details should vary. If a model cannot use a tool or cannot accept a large context, the assistant should say so before the request fails.
Practical follow-up
Provider switching is most useful when the assistant preserves the conversation contract while exposing capability differences. A local model may support a smaller context or fewer tools, so the selector should show those constraints before the request starts.
Keep provider credentials outside the conversation state and make retries provider-aware. A timeout from a cloud API should not automatically replay a potentially expensive request against a second provider without the user understanding the change.