Chukke documentation
Assist and Providers
Configure optional model providers and review document content before each Assist request.
Assist is an optional workflow for running predefined recipes with a model you configure. It is separate from Model-ready, which runs locally.
Enable Assist
Assist is off by default. In Settings → Assist, add at least one model, then turn on Enable Assist. The switch remains unavailable until a model is configured. Supported provider types include:
- local Ollama
- OpenAI-compatible services
- Anthropic Claude
- Google Gemini
Provider availability, privacy, and retention are governed by the service you select.
You can test a connection while adding or editing a model without enabling Assist. You can also keep multiple model configurations and choose the active one. Deleting the last model turns Assist off, but your recipes remain available for the next model you add.
Review every request
Before an Assist action sends document content, Chukke shows:
- the selected provider
- the endpoint
- the model
- the source scope
- the full system instructions and user prompt, including the selected Markdown and recipe or Refine instructions
The endpoint shown is the final provider request URL. Chukke sends the shown message only after you confirm the preview; a configured API key is sent separately when required and is not displayed. You can cancel a pending request before it completes. Test Connection in Settings is separate: it sends a short provider-specific check, and the configured API key when needed, without document content or a payload preview.
For a file opened outside the active workspace, the local interface may show its path so you can recognize it. Provider request labels use the filename, and the canonical external path is not added to the provider payload. The document content and any explicitly selected workspace context appear within the user prompt in the request preview.
API keys
Keys are session-only by default and must be entered again after restarting Chukke. To avoid re-entry, choose Remember in macOS Keychain for a model when saving its key. Chukke restores that key on later launches on this Mac. Keys are never written to Chukke settings, documents, or diagnostics.
Each model shows whether its key is remembered, available for this session, or absent. Forget key removes both the current key and its Keychain item. Deleting a model also removes its remembered key; Reset Preferences removes all Chukke Assist Keychain items. Changing a model’s provider or endpoint requires a new key so the old one cannot be sent to a different destination. If macOS Keychain is unavailable, an entered key can still be used for this session, but Chukke will not say it was remembered.
If a selected model needs a key and none is available, Chukke stops before sending a request and asks you to add one in Assist settings. If a provider rejects an entered key, Chukke cannot tell whether it expired, was revoked, or is otherwise invalid. The request preview keeps the provider’s error visible and offers Open Assist settings. After changing the key, start the action again and review and confirm its payload; Chukke never retries automatically.
Provider requests carrying an API key require an HTTPS endpoint. Credential-free Ollama can use HTTP on your own machine or local network.
Other provider errors also remain in the request preview with an explicit Try again action. A denied-access response can mean the account or model lacks permission; a rate-limit response can mean requests are being limited or the provider quota was reached. Connection and timeout messages distinguish network or endpoint problems from a provider response. Test Connection keeps the same explanations beside its button in Settings.
Rewrite and review results
Rewrite responses open a color-coded diff so you can accept or reject the replacement. Review recipes return feedback without replacing the source document.
Chukke rejects generated text above 2 MiB before it can enter a review or replace document content. For large rewrite and Model-ready comparisons, Chukke keeps the complete original and result in the two review panes but skips the computed line-by-line diff and changed-line totals. This fallback is used when the two inputs exceed 512 KiB combined, either side exceeds 2,000 lines, or their line-count product exceeds 2,000,000. You still review and explicitly accept or reject the full result.
Local and remote providers
A local Ollama service can process requests on your own configured machine or local network. A third-party provider receives the exact content you approve and handles it under its own privacy and retention terms. Chukke does not control a third party’s retention.
Before using a remote provider, review its current privacy policy and do not share content you are not authorized to send.