Providers and local models
Use Anthropic, OpenAI, Google Gemini, any OpenAI-compatible server, or the built-in local model that runs fully offline.
Quap does not include an AI subscription. You choose where the planning model runs: a cloud provider with your own API key, your own server, or a model that runs on your computer.
| Option | Runs on | Needs | Good for |
|---|---|---|---|
| Anthropic | Anthropic's cloud | API key | Best plans for long, multi-step instructions |
| OpenAI | OpenAI's cloud | API key | |
| Google Gemini | Google's cloud | API key | |
| External server (OpenAI-compatible) | Any server that speaks the OpenAI API — for example Ollama or LM Studio on your network, or a company gateway | Endpoint URL, API key optional | Teams with their own models |
| Local (built-in) | Your computer | For now an Apple Silicon Mac with 8 GB+ memory, and a one-time download | Fully offline work, sensitive data |
Set up a cloud provider
Open AI settings
Settings → AI Assistant, or click Open AI settings in the chat.
Pick the provider and model
Choose the Provider and type the Model name exactly as your provider lists it.
Save your key
Paste the API key and click Save key. The key is stored encrypted on this computer (the settings show its last four characters). Remove key deletes it.
Test
Click Test connection. Connection works. means you are ready.
Send key names (schema)
Off by default. When on, the provider also receives the names and types of keys in your document — never their values. Plans become more accurate because the model knows that your document has customer.email and not mail. Leave it off if key names themselves are sensitive.
Use your own server
Choose External server (OpenAI-compatible), enter the Endpoint URL of the server (for example http://localhost:11434/v1 for Ollama), the model name and, if the server needs one, an API key. Test the connection as above.
Use the built-in local model
The local model runs entirely on your computer. Nothing leaves the computer: key names are shared with the local model automatically, values never.
Check your computer
Settings → AI Assistant with Local (built-in) selected shows This computer: Apple M2 · 16 GB memory and which models fit.
Download a model
| Model | Download | Memory needed | License |
|---|---|---|---|
| Qwen3 1.7B | 1.8 GB | 8 GB+ | Apache-2.0 |
| Qwen3 4B | 2.5 GB | 8 GB+ | Apache-2.0 |
| Qwen3 8B | 5.0 GB | 16 GB+ | Apache-2.0 |
Click Download. You can cancel and Resume later; every file is verified with a checksum before use, and a damaged download is deleted.
Start
The first request loads the model, which takes 10–30 seconds (Loading the local model…). After 10 minutes without use it is unloaded to free memory.
Local or cloud?
Local models are smaller than cloud models: they handle common one- and two-step jobs well, while long, unusual instructions plan better in the cloud. You can switch at any time; suggestion bubbles marked ⚡ do not use a model at all.
The built-in model needs an Apple Silicon Mac with at least 8 GB of memory. Intel Macs are too slow for it — use a cloud provider or your own server there.
Errors you may see
| Message | What to do |
|---|---|
| No API key is set for this provider. | Save a key in Settings → AI Assistant |
| The provider rejected the API key. | Check the key and that it has access to the model |
| Rate limited by the provider. | Wait a moment and Try again |
| The provider did not respond in time. / Network error. | Check your connection or server |
| Not enough memory on this computer | Choose a smaller local model |
| Not enough free disk space. | Free space for the download |