Cloud capability
Send requests to the selected compatible online provider and model.
- Broad model choice
- Requires network and configured access
- Prompts/context may leave the device
- Provider rates and API costs may apply
Three execution modes
The right engine for the right job
Each mode makes a different trade-off. Hybrid can compare or combine model outputs, but it is not guaranteed to be faster, more accurate or cheaper.
Coordinated models
Use a primary model with eligible online workers and optionally a local model when enabled.
- Primary provider can synthesize outputs
- Task splitting can be enabled
- More moving parts to configure
- May increase latency and API usage
Inference on your PC
Run a supported model through the local llama.cpp server.
- Less dependence on cloud inference
- Quality and speed depend on model/hardware
- Local setup needs disk/RAM/VRAM
- Web and other online tools can still use network
Bring your provider
Models, endpoints and keys. Configured your way.
The Online/Models area supports provider records, endpoint configuration, API keys, extra keys, model names, model fetching and connection tests. Compatible OpenAI-style chat completion endpoints can be configured alongside the built-in examples.
- ✓Name and configure a provider endpoint.
- ✓Test connectivity and fetch supported model lists where available.
- ✓Set the primary model for hybrid workflows.
- ✓Test local Tool Calling before relying on actions.
PROVIDER CONFIGURATION
Name · Your preferred provider
Endpoint · Compatible chat completions URL
Model · Select or enter exact model ID
API Key · Store locally; protect the app data folder
Test · Validate connection and response
NEXT · Choose mode and send a simple request
Name · Your preferred provider
Endpoint · Compatible chat completions URL
Model · Select or enter exact model ID
API Key · Store locally; protect the app data folder
Test · Validate connection and response
NEXT · Choose mode and send a simple request
The provider list and model capabilities can change. A model that replies to chat may still perform tool calling weakly or not at all.
Local model performance
Hardware shapes the experience
Use the smallest model that meets the task requirements, and verify its ability to call tools. Setup may use a compatible Vulkan runtime with GPU offload; the app can fall back when full offload is not possible.
▦
Size is not everything
Model file size, context length, KV cache, memory pressure and tokens per second all matter. Larger models are not automatically the best fit for every PC.
⌘
Test the action path
Use the Tool Calling test status—such as ok, weak, fail or error—before expecting a model to reliably execute Windows tasks.
⚠
Data handling differs by mode. Online requests send the prompt and assembled context to the configured provider. Hybrid may send parts of a request to multiple providers. Provider API keys are stored in a local JSON configuration and are not encrypted by the current implementation, so protect the app data directory.