A voice assistant that actually listens, not just waits

What it does

Jarvis is a fully local, voice-first AI assistant that runs on your own hardware via Ollama. It listens continuously, responds to its name placed naturally anywhere in speech, and maintains a rolling short-term context so it can chime into ongoing conversations — ask “what do you think?” and it knows what you mean. It persists memories across sessions, redacts sensitive data before saving, and can search the web, read screens, track nutrition, control Chrome, and hook into external tools through MCP.

The interesting bit

The wake-word handling is unusually flexible: an LLM intent judge (gemma4:e2b) classifies whether you were addressing Jarvis, extracting the actual query while filtering out its own speech echoes and stop commands. This lets it participate like a third person rather than forcing rigid command syntax. A knowledge-graph memory auto-organises by topic and surfaces relevant history without manual tagging.

Key highlights

• Embedding-based tool selection claims unlimited MCP integrations without context rot or slowdown
• Built-in dictation mode (hotkey-to-text) runs offline with no subscription
• Automatic tone adaptation — surgical for code, pragmatic for business, etc.
• Web search falls back through DuckDuckGo → Brave → Wikipedia with auto-fetch
• Memory Viewer GUI for inspecting diary entries, knowledge graph, and meal history

Caveats

• Voice-only interface; no text chat yet (#35)
• macOS is the primary development target; Windows/Linux may lag
• Dictation broken on macOS 26+ (Tahoe) due to pynput incompatibility (#172)
• “Stop” commands during its own speech sometimes misclassified as echo (#24)

Verdict

Worth a look if you want a privacy-respecting, always-listening assistant that handles multi-turn conversation naturally. Skip it if you need text input, mobile access, or polished cross-platform parity right now.