Anamanti Core

The brain lives on a Mac in your house.

One Rust service hears every room, decides how to answer, remembers who said what, acts on your behalf and speaks back to the screen that asked. It works out the cheapest way to answer each turn, and it shows you where each step ran.

127.0.0.1:8730 / overview
Anamanti Core overview page
Overview. Each pipeline stage with its health and a Local, Your network or Cloud label, the three connected displays, per-stage latency, and the last few turns tagged with who spoke.

Anatomy of a turn

Wake. Listen. Understand. Decide. Act. Speak.

The display streams audio to the Core over the open Wyoming protocol. Everything after the wake word happens on the Core, and every step streams into the next instead of waiting for it.

Wake word

openWakeWord (“Hey Jarvis”) runs on the display. No audio leaves the room before it fires.

On the display

End of speech

Silero, a small neural voice-activity model, decides when you've stopped talking. It takes under 1 ms per frame.

On this Mac

Speech to text

Your choice of three engines: faster-whisper, whisper.cpp built into the Core, or ElevenLabs Scribe. Swap live from the console.

Mac or cloud

Decide

Fast paths first, then the System-1 classifier, then the full language model. The cheapest step that can answer does.

Mac or cloud

Remember & act

Recall from the memory graph, then up to four rounds of tool calls: lights, timers, music, recipes, search.

Mac + services

Speak

Piper voices each sentence the moment it's complete, so you hear the answer while the rest is still being written.

On this Mac

Want the details on how “Decide” stays fast? See System-1 and the fast paths.

Major features

Everything a household assistant should do, and some things they don't.

Conversation

Follow-ups without the wake word

After every reply the mic reopens: for 10 seconds if the answer ended in a question, 5 otherwise. The conversation continues as long as you keep talking, and ends quietly when you stop.

Conversation

Barge-in

Say the wake word while it's talking and playback stops instantly. The Core abandons the half-written answer and starts listening to you.

People

Knows who's speaking

A voiceprint model identifies the speaker on every turn, on the Mac. Memories are kept per person, shared with the household, or quarantined if the voice is unknown.

Language

Multilingual

Each turn is answered in the language it was spoken in, with a voice to match, and the fast paths understand it too. We're building and testing with English and Brazilian Portuguese right now. The goal is to confirm that the same person's voiceprint matches, and their memories are recalled, whichever language they speak. (In testing.)

Intelligence

Bring your own model

Ollama on your own hardware, Claude (API key or your Claude subscription) or OpenAI. Pick the backend and model in the console; it changes on the next turn, with no restart.

Character

Personalities

“Talk like a 1940s gangster.” A personality changes only how a reply sounds, never what the assistant knows or does. Switch styles by voice and edit them in the console.

Rooms

Many displays, one Core

Every display connects to the same Core, and replies, music screens and camera views go to the room that asked. A display can be pinned to a specific Core, so test and production can run side by side.

Push

Notifications

Each display keeps a channel open to the Core, so the Core can push banners, weather and now-playing to the screens without anyone speaking.

Accuracy

Transcript fixes

Teach it how local names get misheard (“kitchner” → Kitchener) and the transcript is corrected before the model ever sees it.

Voice tools

24 things it can do, not just say.

Tools are offered to the model only when they're configured, so the assistant never claims an ability it doesn't have. See Integrations for the services behind them.

Kitchen

Set, adjust and cancel timers. Find a recipe and walk through it step by step on screen.

Home

Lights, switches, scenes and TV through Home Assistant. A live view of the front-door camera.

Music

Spotify search, play, queue and skip. Internet radio stations. Switch the music screen between full screen and compact.

Out and about

Weather with a full-screen forecast. Place cards with hours and ratings. Directions with live traffic.

Family

Add to the shared shopping list, add calendar events, look up what's on, and send a message or map link to someone's phone.

Knowledge

Real web search. Long-term memory: “remember…”, “forget that”, “who am I?”

The screen

“Make the text bigger.” Close weather, recipe, place or camera views by voice.

Style

Switch personality, or go back to normal.

The console

Nothing is a black box.

The Core serves an admin console that only the Mac itself can open. Each page is designed to fit one screen. Changes apply live, and every option that touches the internet says exactly what it sends.

Next

Meet the face of the house.

The Core is the brain. The Display is what the family sees and talks to every day.