ecce

Design

Made to sit on a shelf and go unnoticed.

It's wool felt and solid ash, about the size of a coffee tin. It fits in the spot where an Echo Dot sits now.

Top-down view of the ash cap with a black touch disc in the center and four tiny microphone holes near the edge.

Macro of the seam between ash wood and felt, with the amber light band glowing.

Cap
Solid ash with a glass touch disc. Tap to talk or stop, and trace the edge for volume.
Mics
Four far-field microphones at the rim, with echo cancellation so it hears you over its own music.
Band
A ring of LEDs between wood and felt, bright enough to read from across the room.
Body
Acoustic wool felt over a printed shell. The sleeve slips off and can be swapped.

The light band

One line of light tells you what it is doing.

Steady amber. It heard its name and is streaming what you say.

Talk to it

Say it the way you'd say it to a person.

You don't have to remember skill names or say “ask so-and-so to”. A language model reads the whole sentence and picks the plugin. Everything below works on the prototype today.

Play some Khruangbin in the kitchen.
Playing Khruangbin on Spotify.
Spotify
Set a pasta timer for nine minutes.
Nine minutes for the pasta, starting now.
Timers · works offline
Tell the bedroom dinner's ready.
The bedroom speaker says it out loud.
Announcements
Add oat milk and lemons to the list.
Added oat milk and lemons.
Grocery list
Do I need a jacket?
Yes, it's 48 and windy until about noon.
Weather · example reply
What's in the news?
Reads the top NPR headlines, or one topic if you ask.
News
Hey Ecce, snooze.
The alarm stops and comes back in nine minutes.
Alarms
Actually, never mind.
It stops mid-sentence and remembers what you were talking about.
Barge-in

The speaker glowing on a dining table during dinner, with people talking in the background.

A fingertip touching the black touch disc on top of the speaker while the band glows amber.

Speed

About a second from your last word to its first.

A pause of two seconds feels broken. ecce streams every step, so it starts talking while the model is still writing the answer.

Transcribes while you talk, so most of the words are in before you stop.
Starts thinking on the draft the moment you pause, and holds any action until the final transcript agrees.
Speaks sentence by sentence, while the rest of the answer is still being written.

Beforefirst build

1.85 s

NowOctober 2026

1.12 s

00.51.0 goal1.52.0 s

Median time from the end of a request to the first word of the reply, measured by the test harness on the bench rig across all test requests. The dashed line is the goal.

Every room

One per room, and each knows where it is.

Ask in the kitchen for music and it plays in the kitchen. Ask it to tell the bedroom something and the bedroom speaker says it.

The speaker in a sunny kitchen beside bread and coffee. — **Kitchen**Timers, the grocery list and music while you cook.

The speaker on a walnut nightstand at dusk, its band glowing dimly beside a lamp. — **Bedroom**Alarms with snooze, and a band that dims at night.

The speaker on a white oak shelf among ceramics and a trailing plant, with its light off. — **Living room**With the band off, it looks like part of the shelf.

Plugins

Anything with an API can be a plugin.

Each ability is a plugin. A plugin is an MCP server, the format thousands of existing tools already use, plus a short file that tells ecce where it runs and how to talk about it.

Timers and alarms

Working

Run on the speaker itself, so they ring even when the internet is down.

Spotify

Working

Play, pause and skip on any Spotify Connect device, Echos included.

Grocery list

Working

One shared list for the house. Add from any room, read it back at the store.

Weather

Working

Open-Meteo, no account needed. Kept fresh so the jacket question needs no lookup.

News

Working

NPR headlines, overall or by topic.

Announcements

Working

Say something out loud in another room.

Home Assistant

Built, not connected

Thermostat, alarm, door sensors and lights, through Home Assistant's own MCP server.

Calendar

Planned

The family calendar, once there is a safe way to connect it.

# plugins/weather/ecce.toml
[plugin]
name        = "weather"
description = "Current weather and the forecast"
mcp         = "stdio:python server.py"
runs_on     = "hub"

# Read before every request, so "do I need
# a jacket?" needs no lookup round trip.
[context]
tools       = ["home_weather"]

[voice]
style       = "brief"

This is the weather plugin's actual manifest. A plugin that can unlock a door or spend money has to get a spoken yes first.

Your model, your data

It thinks in the cloud and listens only in the room.

In the roomSpeakerHears its name, plays the answer

In the houseHome hubLocal plugins, offline basics

In the cloudYour modelThinks, remembers, calls tools

Bring your own brain.

The speaker is mostly microphones and a good driver. The model does the thinking, so when models improve, ecce improves with no new hardware.

ClaudedefaultGPTGeminiOpen models

Use a different model for each person, or for each kind of request.
The brain is one container. Run it on a mini PC at home or on your own server.
Ships as a Home Assistant add-on too.

The speaker in a dim room with its band glowing red because the microphones are switched off.

Red means nobody's listening.

A switch on the cap cuts power to the microphones. Software can't turn them back on.

The wake word is heard on the speaker. Nothing leaves the room before it.
If the TV says something close to “Ecce”, the transcript gives it away and the speaker goes back to sleep without a sound.
Transcripts and memory live in your own database. Export or delete them any time.
It never suggests things to buy.

Finishes

Three felts, two woods.

The felt sleeve and the cap both come off, so you can change the color when you move a speaker to another room. These finishes are concepts.

Three ecce speakers side by side: charcoal felt with ash, oat felt with walnut, and moss green felt with ash.

Charcoal and ashThe original.

Oat and walnutFor lighter rooms.

Moss and ashFor shelves with plants on them.

Specs

What's inside.

Target specs for the first enclosure. The bench rig today runs on a Mac with a USB mic array.

Exploded view of the speaker: ash cap, light diffuser ring, microphone board, compute board, driver and felt shell.

Size: About 100 mm wide, 110 mm tallBetween an Echo Dot and a full-size Echo.
Microphones: 4-mic far-field array, XMOS XVF3800Beamforming and echo cancellation on the chip, so you can interrupt it.
Sound: 3″ full-range driver with a passive radiatorFires down into the table for sound all around the room.
Light: 24-LED band between cap and felt
Controls: Touch disc, touch volume ring, hardware mic switch
Wake words: “Hey Ecce”, “Ecce”, or “Computer”Our own wake word model, still in training. It runs on the speaker, and the cloud transcript double-checks every wake.
Brain: Streaming speech-to-text, an LLM with tools, streaming voiceRuns on a home hub or in the cloud, as a container or a Home Assistant add-on.
Offline: Timers, alarms, volume and “stop”
Power: USB-C or DC barrel

Compared

Why not just an Echo?

Area	Amazon Echo	ecce
The model	Amazon's, and only Amazon's.	Whichever model you choose.
New abilities	Certified skills, usually called by name.	Any MCP server plus a short manifest. You ask, and it finds the plugin.
Conversation	Better with Alexa+, still phrase-bound in places.	Follow-ups, multi-step requests, and interrupting mid-answer.
Your data	Lives in Amazon's cloud.	Lives in your own database.
Offline	Mostly stops.	Timers, alarms, volume and “stop” keep working.
Upsells	Shopping suggestions and “by the way”.	None.

The build

For now it's a prototype in one house.

The first goal is to replace the Echos in our house. If that works, ecce might become a product or an open-source project.

Code on GitHub

Phase 0 · now

Bench rig

Wake word, streaming, barge-in, timers and offline on a Mac. Waiting on the mic array.

Phase 1

Kitchen unit

The first felt enclosure, the light band and touch. It replaces one Echo for a month.

Phase 2

Whole house

One per room, voice recognition for each person, intercom, and a memory you can edit.

Phase 3

Open it up

A custom board, a plugin kit and docs.