Everything that uses a model you have already downloaded works offline. Chat, transcription, speech, image work and the tools all run on your Mac, so pulling the network cable changes nothing about them. What needs a connection is downloading something you do not have yet.
Works with no connection
- Every model already on disk, for every task it supports.
- All the local tools: files, system control, browser automation against a browser on this machine.
- The knowledge base, which ships inside the app and answers offline once its encoder model is on disk, and the in-app help pages.
Needs a connection
- Downloading a model you have not installed.
- Cloud model providers, if you have configured one with your own API key. Those are calls to that provider, made by your machine, using your key.
- Checking for an app update, which happens on its own every six hours when a connection exists and only reads the release list.
This is the architecture, not a fallback
ToolPiper does not call home. There is no telemetry path, no remote configuration fetch and no background sync, so offline is the same code path as online rather than a degraded mode that gets less testing. Prompts, files and all inference stay on your Mac. The one thing that can leave is an anonymous benchmark score, only if you run the benchmark, and it shows you exactly what it sends before it sends anything.
Working on a plane
Download the models you want before you go and nothing else is required. There is no license check on the way to a local model and no first-run activation that has to happen online.
If you use a cloud provider through your own key, that one path stops until you have a connection again. Local models keep working.