Voice runtimeSelf-hosted
Give your agent a voice. Keep it in the house.
You already run an agent. Vulpine lets you talk to it. It hears you and answers simple requests at once. It sends everything else to your agent, then speaks the reply in a voice you chose. It runs on a Mac in your house, with apps for your screens and speakers for your rooms.
Install it See how fast it answers
- Speech in
- on your Mac
- Voice out
- on your Mac
- Simple requests
- answered at once
- Agent
- Pi today, more to come
- Account
- none
VulpiListening
Move over the fox, or tap it. A ripple leaves from its ear.
01 / The turn
Two ways to answer.
A turn is one request and its reply. The gateway is the Vulpine program on your Mac. It answers simple requests itself, at once. It sends everything else to your agent, and only those turns wait.
Is it a verb?
A verb is a request the gateway can do by itself.
-
YesThe gateway does it
0.1 s to its first word
-
NoIt goes to your agent
1.9 s to its first word
Yes, a verb
The gateway does it.
The gateway knows a short, fixed list of requests. It does them itself, and your agent is not asked.
turn off the office lights
The lights go off, and it says what changed.say that again
It repeats its last reply.pull up retro weather on the screen
It shows that page on the screen.never mind
It drops the request. Nothing is sent.start over
The conversation starts again from nothing.
No, anything else
It goes to your agent.
Questions, plans and open requests. The gateway sends your words to your agent, and your agent uses its tools.
any weather warnings?
Your agent looks them up and tells you.play jazz radio in the kitchen
Your agent finds a station and plays it there.what's on my screen
Your agent reads the screen and tells you.what's this error
Your agent reads the error and explains it.I'm in the kitchen
Your agent notes which room you are in.
Each step, its time and its receipt
A receipt is a measurement file saved beside the code. The receipts call your agent the brain, and its first word a token. The commit is the moment the gateway is sure you have finished.
-
AFinishat home
978 ms
from the end of your speech to the turn's commit; both paths wait for it
Speech end to commit on the home Mac: the median of 218 live turns, 29 October to 6 November.
Receipt, 2026-11-06:
brain-ttft-cloud-vs-local-2026-11-06.json -
BMatchat home
8.6 µs
for the gateway to recognise one of its own verbs at the commit
The phrase match the commit runs, per call, over 80,000 calls in an unoptimized build: an upper bound.
Receipt, 2026-11-07:
page-verb-cost-20261107.md -
CSpeakat home
140 ms
from text to the first audio, when the voice has been quiet for up to a minute
The voice worker's time to first audio on the home Mac after 5 seconds to 1 minute of quiet: the median of 457 live requests. It is quicker when it has just spoken and slower after an hour of silence.
Receipt, 2026-11-07:
tts-idle-ttfa-2026-11-07.json -
DThinkaway
1.4 s
from the commit to a hosted brain's first token, the one leg that leaves the house
Commit to first token with a hosted model: the median of 218 live turns, 29 October to 6 November.
Receipt, 2026-11-06:
brain-ttft-cloud-vs-local-2026-11-06.json -
ESpeakat home
441 ms
from the brain's first token to the first audio of the reply
First token to the first audio chunk: the median of the same 218 live turns.
Receipt, 2026-11-06:
brain-ttft-cloud-vs-local-2026-11-06.json
What these numbers do and do not say
Read it honestly. Each step on the agent way is a median over the same live turns. But medians of steps do not add up to a median turn. Traced whole, an agent turn takes 3.0 s from the end of your speech until you hear the reply. One turn in ten waits 3.8 s or more for the agent's first word. No verb turn has been timed whole yet. Some verbs answer with an action or a panel, not a sentence.
The short way is the small one today. In one week of live turns, 14 of 230 were verbs (6 in a hundred). The rest went to your agent. After the commit, a verb answers about 13 times sooner. Your agent does not wait idle. On 117 of 229 turns, the gateway started it before you finished speaking. That is the early start.
The long step is the model. Today the supported agent runtime is Pi, installed on your Mac. More runtimes are to come. The model your agent calls is usually a hosted one, and that call is the one step that leaves the house. A model on your own Mac has no live turn timed, so no number is claimed for it. How the install guide sets Pi up
Every step has a budget. A change that adds time to one step must name the step that gives the time back. That rule is the turn budget.
Where the decision model sits
Beside the turn, not on it. The decision model is a small model on the home Mac. The gateway asks it three questions while you speak: was that for Vulpine, have you finished, and which verb. Its answer takes 682 ms, and the turn never waits for it. Today the gateway matches verbs from its own list. The model can do one thing: drop a follow-up that was not meant for Vulpine. That is the drop gate. Its next job is to send more turns down the short way. It is being measured for that job under the decision contract.
02 / One hub
One hub. Six lanes.
Most assistants are a chain. One service calls the next, and each hop adds delay. Vulpine puts one gateway in the middle, on a computer in your house. Each lane is one job, such as listening or speaking. Watch four events go round the hub. Each one takes its own route.
A verb turn
You say “Turn off the office lights.”
- Listen You speak to a room speaker. Your words become text.
- Turn-taking You stop. The gateway is sure you have finished. 978 ms
- Decide It is a verb the gateway knows. Your agent is not asked. 8.6 µs
- Act The gateway runs it through the home adapter.
- Speak It says what changed. 140 ms
An agent turn
You say “What is the weather tomorrow?”
- Listen You ask from the desktop app. Your words become text.
- Think, early Your words so far hold steady. The gateway starts your agent early and holds its reply back.
- Turn-taking You stop. The gateway is sure you have finished. 978 ms
- Decide No verb matches. The turn goes to your agent.
- Think Your agent's first word arrives. 1.4 s
- Speak The gateway speaks the reply. 441 ms
A camera event
In the house Someone is at the door.
- Act A camera sees someone. The event comes in through an adapter.
- Act Room presence says which room you are in.
- Speak The announcement plays on that room's speaker.
A fleet task
You say “Run the tests, and write the docs.”
- Listen You hand out work. Your words become text.
- Turn-taking You stop. The gateway is sure you have finished.
- Decide No verb matches. The turn goes to your agent.
- Think Your agent plans the work.
- Act The fleet grows. Coding agents start on your machines.
- Act They finish, and their reports come back.
- Speak It tells you what is done.
Screens and speakers
- macOS the desktop app
- Web any browser
- Speaker a room speaker
- iOS in development
The house
- Rooms speakers, presence
- Home verbs out, events in
- Cameras events in
- Fleet agents on your machines
The six lanes and the hardpoints on each
-
Listen
The gateway hears the wake word. Then your speech becomes text, a few words at a time.
wakesttcounted inspeech_end_to_commit, on your machines -
Turn-taking
Have you finished? The gateway waits for the pause after your request. Then it is sure.
answered by
decision, mounted on Decide counted inspeech_end_to_commit, on your machines -
Decide
Is it a verb the gateway runs itself? Then your agent is not asked. Anything else goes to your agent.
decisioncounted inspeech_end_to_commit, on your machines -
Think
Your agent, with its memory and what it can look up. Today the supported agent runtime is Pi, installed on your Mac. More runtimes are to come. It gets the text of a turn, never the audio.
brainmemoryretrievalcounted incommit_to_first_text, on your Mac (its model can be hosted) -
Act
Verbs go out to the house through adapters. Events come in the same way. Each action returns a result.
roomcounted intool_round, adapters beside the hub -
Speak
The reply becomes audio in the voice you chose.
ttsvoicecounted infirst_text_to_first_audio, on your machines
A hardpoint is a named slot for one part you can swap. Nine hardpoints mount on the hub's lanes.
The other two,
avatar and render, mount on each screen. The loop is
slowed down so you can read it. It is not to scale.
03 / Within reach
The house within reach.
Rooms, lamps, cameras and your other machines reach the hub through adapters. An adapter is a small program that connects one kind of device. Screens and speakers connect to the hub directly. Your agent runs on your Mac. The model it calls is the one part that can sit outside the house, on the dashed line.
04 / Hardpoints
Every part is a slot.
To start, bring a Mac with Apple silicon and the agent you already run. Today the supported agent runtime is Pi, installed on that Mac. More runtimes are to come.
The install downloads the rest: speech recognition (about 630 MB) and the voice (about 1.3 GB). There is no server to stand up and nothing else to bring. What an install needs
A hardpoint is a named slot for one part. You can swap the part that is included for your own. 2 slots take a swap today: speech to text and text to speech. The other slots have their names, and their swaps come next.
-
01Swap works today
sttSpeech to text
- Included
- Parakeet, on your Mac. Downloaded at install.
- Or swap in
- your own speech server on your Mac. The catalog has adapters for two kinds of server.
-
02Swap works today
ttsText to speech
- Included
- Qwen3-TTS, on your Mac. Downloaded at install.
- Or swap in
- a Wyoming text-to-speech server on your Mac. The catalog has one for Kokoro.
-
03Swap is next
wakeThe wake word
- Included
- A spoken name, found in the text of what you said. A room speaker hears its own wake word.
- Later, swap in
- your own wake word detector
-
04Swap is next
decisionThe decision model
- Included
- Nothing needed. The decision model is optional and off by default.
- Later, swap in
- any model server that answers the decision contract
-
05Swap is next
brainThe agent that answers
- Included
- Not included. Bring your agent. Today the supported agent runtime is Pi, installed on your Mac.
- Later, swap in
- more agent runtimes, as they are added
-
06Swap is next
voiceThe voice you hear
- Included
- The voices that come with a persona.
- Later, swap in
- a sample of your own voice, with consent
-
07Swap is next
avatarA face, or a fox
- Included
- Vulpi, the 2D fox.
- Later, swap in
- your own ready asset, such as a VRM or Live2D rig
-
08Swap is next
renderThe visual panel
- Included
- The conversation window and its panels.
- Later, swap in
- your own frontend that says which panels it can draw
-
09Swap is next
roomRooms and the house
- Included
- Room and home adapters, and the room speaker's bridge. All optional.
- Later, swap in
- your own adapter program or a shell script
-
10Swap is next
memoryWhat it remembers
- Included
- Its memory of what you told it, kept on your Mac.
- Later, swap in
- your own memory store
-
11Swap is next
retrievalWhat it can look up
- Included
- A search index, kept on your Mac.
- Later, swap in
- your own search provider that cites its sources
-
Key
This table follows a design decision on record.
05 / Surfaces
Where you talk to it.
-
macOS, the desktop app
Vulpine.app
A face at the edge of the screen that listens, answers and shows its work.
Its body and its face are separate parts. You can swap one and keep the other.
-
Web, any browser
Vulpine in a browser
A full-screen call in a browser tab, with frames, portraits and a running transcript.
Each frame shows the same conversation, so a second screen can be the voice's face.
-
Rooms, a puck on a shelf
A room speaker
A small speaker and microphone for a room, with the wake word heard on the device.
Its ring of lights shows when it is listening, thinking or speaking.
-
iOS, in development
Vulpine on iPhone
Hold a button to talk, or type. It is the same conversation, in your pocket.
It pairs with the gateway at home through an invitation code.
-
macOS, the host
Vulpine Gateway.app
The app that runs the gateway. One window shows sessions, speech, your agent, the house and the approvals.
It reads the gateway over a local connection that only your account can open. It never changes settings itself.
06 / Not included
What is not in the box.
-
No cloud account
Nothing to sign up for. It runs on a computer in your house.
-
No always-on upload
Your machines process the audio. Your agent gets the text of a request, never a live microphone stream.
-
No advertising
Nothing is sold, and nobody builds a profile of you.
07 / Install
Install it in four steps.
There is no account and no price. You host Vulpine yourself, and nothing is sold. The source is not public yet. You can check what you need and read the install guide now.
Needed
A Mac with Apple silicon
It needs 16 GB of memory, and about 7 GB free for the install, plus room to build the gateway.
Needed
Pi, installed on that Mac
Today Pi is the supported agent runtime. More runtimes are to come.
Optional
A speaker for a room
A small puck with a microphone. It hears the wake word itself.
Optional
A screen for a face
Vulpine.app on the desktop, or a browser tab.
Check the Mac
Check memory, disk and macOS before you install anything. Before you start
Install the tools
By hand, once: Apple's command line tools, mise, cmake and ffmpeg, and Pi.
Run the install
One command builds the gateway, writes its settings and downloads the voice. Speech recognition and the permissions helper are two more steps by hand today. The install sequence
Let the doctor finish
The doctor is the last step. It checks every part and names the fix for anything missing. The install guide
08 / Questions
Common questions.
What do I need to run it?
A Mac with Apple silicon in your house, and an agent. Today the supported agent runtime is Pi, installed on that Mac. More runtimes are to come. The install guide starts with a few tools you add by hand. Then one command builds the gateway, writes its settings and downloads the voice, and two short steps by hand follow it. The guide ends with a doctor that checks every part and names the fix for anything missing. Install
Does my audio leave the house?
No. Your own machine turns your speech into text, and it makes the voice too. Household audio never leaves your machines. A check on every change to the source refuses a command that would copy audio anywhere else. Why speech stays home
What does my agent see?
The text of a turn that needs it, never the audio. Your agent runs on your Mac. Today that is Pi, installed there, with more runtimes to come. The model it calls is usually a hosted one. Speech recognition and the voice stay on your machine. What the model sees
How does it know I am talking to it?
Today, by a wake word and the pause after your request. A small decision model on your Mac is also asked three things: were you talking to it, did you finish, and what did you mean. It can already drop a follow-up that was not meant for it. The rest of what it answers is recorded and routes nothing yet. The decision contract
Does it know who is speaking?
It can compare a voice with the voice you enrolled. It only records the result. The gate that would act on the result is off by default, so it refuses nobody. Voice commands
Does it read my mood?
Only if you opt in. Every part is off by default. When it is on, signs that a turn went wrong can make the next reply shorter or more careful. It never says how you sound, and its notes stay on the home machine. The mood decision
Where do the numbers on this page come from?
From measurements saved beside the code. The page reads them when it is built. The build fails if someone types a number in by hand. How the site is built
Can I see what it is doing?
Yes. Vulpine Gateway.app shows sessions, speech, your agent, the house and the approvals. It reads the gateway over a local connection that only your account can open. The Gateway window