Offline Agent
The Offline Agent is Beta on iPhone and iPad. On Android it is labelled Experimental in Deskless 3.0 (in review on Google Play): the Gemini Nano model often misses requirements and has not yet reached the quality of Deskless’s other agents, so check and review every proposal. Android 2.1 labels it Beta.
Offline Agent proposes small project edits using the downloaded device model. It is free, needs no provider API key and has no cloud fallback. It is separate from the Pro cloud Agent Mode.
Check model availability
| Platform | Requirements |
|---|---|
| iPhone and iPad | Supported Apple Intelligence hardware, Apple Intelligence enabled, and the system model downloaded |
| Android | A supported Gemini Nano/AICore device with the required model and structured-output capability |
An app update alone does not make an unsupported device eligible. Android offers Prepare model in the Offline Agent screen; preparation can download model data and needs connectivity. Prepare the model and project before working offline. If readiness fails, follow the reported device/model limitation rather than entering a cloud key here.
Propose a small change
- Open and configure a project in Project Tools, including tests that check the behavior you care about.
- Choose Offline agent, marked Beta on Apple devices or Experimental on Android 3.0.
- Describe one small fix or refactor and select one to three source files. Optional focus symbols help target a larger file.
- Choose Generate local proposal on Apple devices or Generate proposal on Android.
- Review each proposed change and select the edits you want to keep.
- Choose Check selected edits. Changing the selection invalidates the earlier check results.
- Inspect the test and diagnostic results, then apply the selected edits if appropriate. Save separately to write the drafts to the project.
The model can make up to three attempts. It may use excerpts from larger files, while checks still use the captured project and configured tests. Tests, configuration, dependency locks and dependencies stay outside the model’s editable selection.
Read the results carefully
Recorded checks passed means the recorded checks for that exact candidate passed. It is not proof that every requirement is satisfied. If checks fail or tests are missing, the screen says so; the explicit apply action preserves that warning. An empty test set is not a passing test suite.
Models can miss requirements, produce invalid edits or make no useful change. The Beta and Experimental labels are intentional; these docs do not claim universal reliability across devices or model versions. Keep tasks small and review the resulting source and Git diff.
Stop and resume
Stop, closing the agent, or leaving the foreground cancels active work. A saved proposal can be recovered when you reopen the same unchanged project; run fresh checks before applying it. If the project changes, capture a fresh proposal instead of applying stale edits.
Local processing keeps this workflow on the device. Choosing cloud Agent Mode or another cloud AI action is a separate workflow with its own content permission and provider terms.