Voice

tap to unlock

Nothing said yet

Settings

Brain

One brain at a time for the whole app, and as many saved as you like. Any OpenAI-compatible /v1 endpoint — LM Studio, llama.cpp, Ollama, vLLM, OpenRouter, OpenCode. An agent chooses its instructions, language, voice and folder; which model is running is a property of the machine.

Saved brains

Base URL API key

Model

OpenCode lists its whole catalogue, including models it serves on OpenAI's /responses and Anthropic's /messages — which this app does not speak, so those refuse here. On the endpoint that works: GLM, Kimi, DeepSeek, MiMo, Space Bunny, and the free LongCat and Step models.

Temperature Max tokens

Max tokens 0 means no limit — the local server decides. A ceiling here is how long answers get cut off mid-sentence, and a reasoning model can spend the whole budget thinking and return nothing at all.

Vision model — only for looking at your screen

A 12B brain and a vision model rarely both fit in memory, so this is usually a choice about which model the local server has loaded.

Can the brain above see images?

Work it out guesses from the model's name, and a name is not a promise — a capable brain has been misread as blind before. If looking at your screen does nothing, set this to Yes. With No, the vision model above describes the screen instead.

Recognition engine

checking

Where speech is transcribed. Which languages each agent listens for is an agent setting below.

Backend URL

Which recogniser the ear runs is a property of each agent, not of the server — a conversation and a meeting want different ones. It is set below, next to the language, because the two are the same decision: how to hear this agent.

Synthesis engine

checking

Where the reply is spoken. Each agent picks its own provider and voice below; this is the server both of them talk to.

Voice backend URL

Used by the PocketTTS provider. Piper runs in-process and ignores this.

Microphone

Speak normally. If this bar does not move, the app is not hearing you — that is a microphone problem, not a tuning problem.

—

Trigger level 0.012

Turn taking

How long a pause ends your turn, and how much audio is kept before it.

End-of-turn pause 700 ms Minimum speech 300 ms Pre-roll 400 ms

Allow interruption Hands-free

Only your voice

checking

Records a few seconds of your voice once, then only lets a voice close to it interrupt a reply. Everything else is ignored however loud it is — the tap, the dishwasher, and somebody talking on a video. This is a convenience, not a lock: anybody who sounds like you can still interrupt. It also never blocks you when it cannot run, so a failure here costs you nothing.

Only my voice may interrupt

Running commands

ask

What the agent may change. In workspace write it has free rein in the project folder and asks before touching anything outside it. Reading is always allowed in every mode.

Some things are refused in every mode — deleting your home folder, sudo, formatting a disk, piping a download into a shell.

Conversation

The agent remembers the last 24 messages. Clearing them starts fresh — it does not change any of the settings above.

Typing

The message bar lets you type to the agent instead of talking. Turn it off and it disappears completely — Cmd+K stops doing anything.

Appearance

System follows your Mac's appearance and switches on its own.

Sphere

Three answers to the same question, and all of them are kept in step with everything else: the colour of the state, your microphone's level, the sphere's small size. Click the particles while the app is idle and they scatter.

Interface sounds

Short synthesised cues, not recordings: a chime when the microphone opens, a tick when a tool runs, a settle when the reply ends, a quiet click on buttons, and a rising pair of notes when a page is drawn. Turn them off and the app is silent apart from speech.

Pages on the canvas

When an answer is a shape rather than a sentence — a table, a plan, a comparison, a set of steps — the agent draws it on the canvas beside the conversation and says one line about it, instead of reading the whole thing out. Switch this off and it will only ever speak.

Draw pictures

Ask for a picture — "draw me a fox in a suit" — and it is generated on this Mac and shown on the canvas. About thirty seconds each, and a gigabyte of memory while it runs. Switch this off and the agent will not offer to draw at all.

Reading aloud

checking

Select text anywhere on your Mac — a browser, a PDF, an email — press ⌃⌥R, and your agent reads it in its own voice. Press it again to stop. This is a separate channel: the text is read, not sent to the model, so nothing about it is shortened or summarised.

Read what I have selected

—

Using the computer

checking

Kouto gets its own pointer, which it uses to point at things on screen while you talk, and to click, type, scroll and drag when you ask it to. Your own pointer is never moved — the one it draws is its own, and your arrow stays exactly where you left it.

It acts only when you ask, refuses password fields, refuses shortcuts that log out, lock the screen or force-quit, and stops instantly on ⌃⌥S. Anything on screen — a web page, an email, a document — is information to it and never an instruction.

Let Kouto use the computer

—

Pointer size — 72px

The cursor is a teardrop whose rim and point are fractions of this number, so these are the four sizes it is designed at rather than a range: 48 when you want it out of the way, 72 by default, 96 and 120 when you want to be sure where it is. Below 48 the point is too small to indicate a single word accurately; above 120 the body starts behaving like a small window and covers the text it is pointing at.

Diagnostics

Transcribe: —
First token: —
First audio: —
Turn total: —

Who this agent is

Everything below belongs to the agent selected above — its instructions, its folder, the language you speak to it, the voice it answers in, and its automations. Switch agent and all of it changes with you.

Name Instructions for this agent

Project

none

Attach a folder and the agent can list, read and search the files inside it. It is confined to that folder and cannot reach anything outside.

Folder

Ears — this agent

Leave the language on Auto-detect and you can switch between languages mid-conversation — everything else forces Whisper to translate into that language instead of transcribing.

Spoken language Languages I speak

Auto-detect can misfire on short or quiet audio. If it hears something outside this list, the utterance is transcribed again forcing the first language. Leave empty to always trust the detector.

Recogniser model

Bigger hears better and takes longer. Measured on this Mac, four seconds of audio: base 0.5s, small 1.2s, medium.en 2.5s. small is the one to talk to; medium.en is what meetings use, because nobody is waiting on them.

parakeet runs on this Mac instead of the voice server: about 0.35 seconds for seven seconds of speech, and your audio never leaves the machine. If it cannot start for any reason, transcription falls back to the server rather than stopping.

Voice — this agent

Provider

Voice

Match the voice to the language I speak

With this on, speaking German switches to a German voice automatically and speaking English switches back. Turn it off to always use the voice above.

Automations

Times when this agent speaks first. It never interrupts a reply, and if it comes due while the app is shut you are asked rather than shouted at. The instruction below is an ordinary request: it can use any tool.

Pause every automation

Name When

Minutes apart

What this is

What it should say or do

Links — optional, read fresh every time it fires

With links, it reads them when it fires and answers from what they say — so an automation on a news page gets today's page every time. Refer to them in the instruction above. Without links it will search or use tools on its own if you ask it to.

Listen for a reply afterwards Land as a notification instead of saying it

On, it says nothing when it comes due. A bar appears over the sphere saying it is waiting, with a button to hear it — and it waits until you deal with it. Right for anything that would otherwise talk to an empty room at six in the morning. Off, it speaks the moment it is due.

On, it opens the microphone when it has finished speaking so you can answer straight away. Off, it says its piece and stops — right for anything that is a briefing rather than a question.

Podcast

Two voices working through something you have actually read. Give it links and it reads those and nothing else; give it only an instruction and it searches. Then it writes a script in which every number and name comes from what it read, and the two hosts talk it through.

Which agent writes it

The agent supplies the instructions, the language and the memory. An episode made by your German agent comes out in German; one made by an agent whose notes are about a particular subject will know that subject.

What should they talk about

Links to read — only these, when you give any

With links it reads those and nothing else. With none, it searches the web for the instruction above instead.

First host Second host

How long should the episode be

Wait as a notification instead of playing straight away

On, a scheduled episode is made and then waits: nothing is played, and a bar appears over the sphere saying it is ready. Opening it plays the episode that was made. Off, it starts playing as soon as it is finished. Either way it is the same episode — you are choosing when you hear it, not what it says.

Radio

Live radio from a worldwide directory, grouped into four regions. Say “turn on the radio” and it starts on your region’s default station; the list is one tap away and changes it from there. Nothing is recorded and nothing is summarised — it is simply playing, like a radio.

Radio

Off means the tool is not offered to the agent and “turn on the radio” does nothing. The list, the player and the microphone button are all unchanged — this is the switch for whether the app can start a station at all.

Where “turn on the radio” starts

Which region the default station comes from. The picker still reaches all four — this is only what it opens on. Berlin has its own city list; Germany, India and the United States are national, including each country’s news services.

Record now

not recording

Starts a recording straight away, aimed at whatever is using the microphone right now. Watching, below, is for being asked automatically — this is for when you already know. The transcript is written when the recording ends.

Recording meetings

…

A call can be recorded and turned into a transcript, entirely on this Mac. Your microphone is you and the meeting application's own audio is the other person, so who said what is never a guess. Watching never records anything on its own — it only decides when to ask, and an application you have said no to is not asked about again.

Watch for meetings

Which applications are worth asking about. A switch that is off means Kouto does not even look at that application; it does not mean the app is blocked from running.

What to record

Capture

On speakers, the other person's voice reaches your microphone as well as the tap, so their words arrive twice. Kouto drops the microphone's copy — it is the echo of a loudspeaker, and keeping it would label their sentence as yours. Headphones avoid the problem entirely.

Language Recogniser model

A meeting can afford a larger, slower, more accurate model than a conversation can, because nobody is waiting while it runs. Empty means the same model the rest of the app uses.

Keeping recordings

Keep the audio

Two channels of an hour's audio is roughly 230 MB, which is why this is a setting rather than a footnote. The transcript is never deleted here — only the audio it was made from.

Past meetings

…

Every meeting is also written into your conversations, so the transcript is searchable and the agent can refer to it later. This list is the recording itself: play it back, rename who you were talking to, or delete it.

The shelf

…

PDFs you have added. Press play and your agent reads it aloud from where you left off, in its own voice, until you stop it. Nothing is summarised and nothing is shortened — the text is read as written.

Reading voice

Voice

Used only for reading books. Leaving it on Same as my agent means the voice you talk to reads to you as well; pick another and it will not change how your agent speaks anywhere else.

Pacing

How a book is read aloud. The reading is spoken in pieces, and each piece ends with its own fall and its own pause — so where the pieces are cut is what a reading sounds like.

Piece length 500

How much text is spoken before the reading takes a breath. Sentences are kept whole and several travel together up to this length, so a longer piece means fewer breaks and a reading that flows; a shorter one means more breaks and a choppier, more deliberate read. Below about 150 a piece is shorter than the sentence it is cut from, and the voice starts stopping mid-sentence. Above about 600 there is nothing left to gain: the speech engine cannot say more than that in one breath either, so it splits the piece itself — at sentence ends, so it is not a fault, only a limit.

Pause between pieces Auto

The silence after each piece. Auto is the engine's own — about a fifth of a second, and it lengthens the silence by itself after a very short piece. Left of Auto is tighter than that, right of it is more air.

Takes effect from the next piece of the reading, including one already playing. There is nothing to restart.