Eyra evaluates privacy per action. The runtime reports whether an action leaves the machine, where it goes, whether it is allowed, and whether it required opt-in.

​
Decision model

evaluate_privacy_boundary() handles three main paths:
Action prefixDestinationRequires opt-in
model.Local or remote model providerOnly for remote provider
network.Remote network tool targetYes
realtime.OpenAI RealtimeYes
Unknown local runtime actions are treated as local by themselves. Connector privacy is reported separately from connector manifests through ConnectorRegistry.capability_snapshot(). Each connector declares data classes, destination, local/remote status, risk tier, capabilities, approval requirement, and acceptance state.

​
Capability snapshot

build_capability_snapshot() reports:
  • Main, worker, and vision models.
  • Whether the provider is local.
  • Local Whisper status.
  • Screen capture status.
  • Filesystem roots.
  • Network, OS, browser, MCP, agent, and Web gates.
  • Connector config status, acceptance state, privacy declaration, and remote/local boundary.
  • Remote paths.
  • Local job store path.
  • Common privacy boundary decisions.

​
Defaults

Default behavior:
  • Provider is local Ollama.
  • Network tools are disabled.
  • Realtime voice is disabled.
  • Realtime tools are disabled.
  • Agent and MCP bridges are disabled.
  • Web UI is disabled.
  • Connectors are disabled.
  • No telemetry or analytics.
Ask Eyra: 
Are you local right now?
What would leave my machine?
What can you control?