Local AI and Network Boundaries
Pitchblend is local-first, not automatically air-gapped. Its normal workspace, local model inference, chats, uploaded-file text, memories, settings, and workflows live in the local Geist data store. You decide when to add features that require an internet or peer connection.
What stays on this computer
With a local model selected for normal text Chat:
- model inference runs through the runtime bundled for your platform;
- conversations and private folders are stored locally;
- uploaded files and their extracted text are stored locally;
- profile, chat-only, and folder memories are stored locally;
- workflow definitions and run results are stored locally.
The desktop wrapper treats the embedded workspace as untrusted. Account tokens, raw license material, protected-storage details, and Memory Sync keys remain in the Electron main process or its isolated sync worker. The workspace receives only the narrow summaries and actions needed to draw its controls.
Actions that use a network
| Action | Why a connection is used |
|---|---|
| Download a managed model | Retrieves the pinned model artifact and verifies it locally |
| Sign in or check a license | Contacts Pitchblend's identity and licensing service |
| Use an online model | Sends the model request to the chosen provider |
| Run web search | Sends the search query to the configured public search service |
| Generate an image | Uses the configured online image provider in the current build |
| Use voice | The shipped composer sends the transcript and prior chat history to its online model provider; first use can also download speech-model assets |
| Set up Memory Sync | Uses the cloud for entitlement, enrollment, presence, and signaling; memory data travels directly between trusted devices |
Tools and voice change the boundary
Tool availability depends on the active model and backend. The curated
Windows/Linux Qwen model supports native tool calls, the current curated macOS
MLX model does not advertise them, and capable online models can use tools.
Web search and online image generation use network services. Document search
reads local uploaded text and supplies matching excerpts to the selected
model; those excerpts therefore leave this computer when the selected model is
online. Review tool arguments and status cards in the transcript.
A local chip still does not describe the separate voice path. Voice uses
an online agent in v1.0 even when normal Chat is set to a local model. Its
model request includes the recognized phrase and stored prior history for that
chat session. MMS/Sesame speech processing is local after its required assets
are present, but first initialization can download those assets.
No runtime plugin marketplace
Pitchblend's built-in integration code is compiled into the reviewed application and checked before injection. The shipped desktop application does not download, discover, or install arbitrary plugins, and it has no plugin marketplace.
For the separate rules that apply to cross-device memory, see Peer-to-peer memory sync. For provider-specific privacy, see Online model providers.