JOSIE-Voice · Talk to your Mac

01 · Two ways to use it

A conversation.Not a command line.

Switch between Voice and Chat with one tap. It's the same conversation either way: start a call, hang up, keep typing, and JOSIE remembers it all.

THE REAL THING

Six minutes.
One uninterrupted call.

A real JOSIE-Voice session on a MacBook Pro M4 Pro. Watch a conversation turn into action, with more than 100 tool calls along the way.

No cutsNo speedupsReal tool calls

Qwen-3 0.6B TTS / ASR · 4-bit
DeepSeek 4.1 Flash via Ollama Cloud

Watch the full demo on X ↗

Here is a 6 minute call within JOSIE-Voice on a MacBook Pro M4 Pro using: Qwen-3 0.6B TTS/ASR 4bit and Deepseek 4.1 Flash on ollama cloud. it does 100+ Tool calls with agentic work in the real world. No cuts or speedup's. https://t.co/xlkTtYk3oQ
— Gökdeniz Gülmez (@ActuallyIsaak) October 18, 2026

Show or hide the transcript mid-call. (mockup)

The sky follows the time of day, stars at night. (mockup)

Clips and screenshots marked "mockup" are rendered from the app's source with demo content, not screen recordings. See every screen ↓

02 · Gets things done

Less to do.More to your day.

JOSIE works with the Apple apps you already use. Ask out loud and it really happens: the reminder is in Reminders, the event is in your calendar, the message is waiting in Messages.

Reminders
Remind me to bring a jacket at 7:30.
Added to Reminders, with the time set.
Calendar
What's on tomorrow? Put lunch with Sam on Friday at noon.
Reads your day and adds the event.
Maps
Find a pharmacy near me that's still open.
Looks up places and directions in Maps.
Messages
Text Alex that I'm running ten minutes late.
Writes the draft in Messages. You tap Send.

A voice call where JOSIE answers a weather question, then runs an add reminder tool call and confirms the reminder in the transcript (mockup) — Mid-call, JOSIE checks the weather, then adds the reminder for you. (mockup)

Each app is off until you turn it on
macOS asks your permission first
Nothing is sent without your tap

03 · Features

Small presence.So many possibilities.

01

Real conversation

Knows when you've finished talking. Interrupt Josie just by speaking.

02

Knows your voice

Learns you from three short phrases and tells people apart.

03

Voices you choose

Pick a preset, clone a voice from a short recording, or describe a new one in words.

04

Private by design

Hearing, voice recognition and speech run on your Mac. Choose a local language model or connect your preferred provider.

05

Takes real action

Adds reminders and events, finds places in Maps and drafts Messages (you tap Send). Each is off until you turn it on.

06

Sees what you show it

Turn on the camera and ask about what it sees. Needs an image-capable model.

07

Adapts to the room

Volume follows background noise, with echo cancellation.

08

Living sky

The background follows the real time of day, under Apple's Liquid Glass design.

09

Personas

Talk to JOSIE, JARVIS, HAL 9000 or a character you write, each with its own prompt, voice and orb colour.

10

Your iPhone, your Mac

In the iPhone preview, pick your Mac as the provider, at home or over Tailscale. The call runs on your Mac, and you can show Josie a photo.

11

Choose how it hears

Speech recognition with Qwen3-ASR, NVIDIA Parakeet, Voxtral Realtime or Apple’s on-device recognizer. All on your Mac.

12

Its own chat prompt

Write separate instructions for Chat and Voice. Muted in a call? Typed messages still get a text reply.

Your documents and past conversations as a 3D map in the Library. (mockup)

Calls, mute and an optional local API, right from the menu bar. (mockup)

Watch every request to the local API flow by, from apps, scripts and your iPhone. (mockup)

04 · Under the hood

From your voice.Back to you.

A few carefully connected stages make a conversation feel natural. Here’s how the Mac listens, works out what comes next, and speaks back.

01 / LISTEN
Find the words.
Echo handling separates playback from your voice. Voice activity and turn detection decide when to listen, then speech recognition turns audio into text.
02 / THINK & ACT
Work through the request.
Your chosen language model answers, retrieves context, or uses enabled tools. Agent mode can take several steps before returning a reply.
03 / SPEAK
Keep it flowing.
Text is split into speakable chunks and synthesized into audio. When you interrupt, JOSIE can stop playback and keep the conversation in step.

Excalidraw Mac voice pipeline: microphone and echo gate, voice and turn detection, speech recognition, language model and tools, speech synthesis and playback, with a return path for interruptions — The Mac call path, with echo handling, speaker recognition and interruption control. Your language model can run locally or through your chosen external endpoint.

TOOLS & CONTEXTHow a request reaches the right tool

The tool registry connects the model to built-in tools, documents, Apple apps and MCP servers. Your enabled tools and access depths determine what is available; actions that require confirmation wait for your tap.

Excalidraw tool architecture: callers feed a registry filtered by access policy, which routes to built-in tools, knowledge search, Apple apps and MCP servers, with confirmation handling — Tool routing, retrieval and access policy. Open full size ↗

IPHONE ↔ MACA call that reaches back to your Mac

In the iPhone and iPad preview, your Mac can act as the voice provider. Audio travels to its pipeline and the reply streams back, with tool calls and interruption events kept in the same session.

Excalidraw sequence diagram showing phone-to-Mac model discovery, connection setup, streamed audio, tool results, interruption, image input and hang-up — The phone-to-Mac call sequence over your network or Tailscale. Open full size ↗

Original Excalidraw architecture sketches. Model examples and limits reflect the diagram revision; available engines and Agent mode settings are covered below.

05 · Make it yours

Make JOSIEyour own.

Choose what powers JOSIE, what it can access, and how it works with you.

01 / THE BRAIN

A mind of
your choosing.

Keep the language model on your Mac with Ollama, connect an OpenAI-compatible service, or choose Apple Intelligence. Set the endpoint and model that fit the way you work.

Local or cloudCustom endpointsSpeech engines

Explore the controls

Pick your backend
Choose an address and model, including a custom model ID. Image understanding depends on the model you select.
Tune how it thinks
Set Full or Light prompts separately for Voice and Chat, with reasoning controls where the backend supports them.
Choose how it hears and speaks
Choose voices from Qwen3-TTS, Kokoro, Pocket TTS and more; recognition from Qwen3-ASR, Parakeet, Voxtral Realtime or Apple. Set supported sizes and 4-, 6- or 8-bit precisions on your Mac.

Models & endpoints showing backend, prompt size and Agent mode controls (mockup) — Models & endpoints · Interface mockup. Click to explore.

02 / THE CONNECTIONS

Connected.
On your terms.

Connect the Apple apps you use, turn on web search, and add MCP servers for tools beyond the built-in collection.

Apple appsAccess controlsMCP servers

Explore the controls

Choose what is available
Enable tool groups individually. Apple app integrations start off until you choose to enable them.
Set the depth of access
Use Read, Add or Full where supported: from looking things up to creating, editing or deleting.
Extend with MCP
Bring additional tools into JOSIE through your configured MCP servers. Specialist agents can use a narrower selection.

Tools & MCP with built-in switches and Apple app access depth controls (mockup) — Tools & MCP · Interface mockup. Click to explore.

03 / THE CONTEXT

Pick up
the thread.

Let JOSIE remember useful facts, find something in an earlier conversation, or pull the right passage from your documents.

Personal memoryDocument searchPast conversations

Explore the controls

Remember, retrieve, forget
Control the Memory tool and its access depth. Ask JOSIE to remember a fact, recall it later, or forget it.
Search your own context
Enable document search and past-conversation search independently, with optional automatic context retrieval.
Tune what comes back
Adjust chunk size, overlap, passage counts, recent turns and similarity thresholds in Documents & search.

Documents & search retrieval controls for chunks, overlap, passages and similarity (mockup) — Memory & knowledge · Interface mockup. Click to explore.

04 / THE TEAM

A little help.
A lot more done.

Turn on Agent mode for work that takes several steps. JOSIE can search, read, check and hand a focused task to a specialist before answering.

Specialist agentsModel poolsBackground work

Explore the controls

Make specialists your own
Give each agent instructions, allowed tools, MCP servers, skills and a step limit. Call one in with @A.name.
Match the model to the task
Build a pool of up to four backend models with capability notes and supported reasoning depths. JOSIE picks per task.
Keep the conversation moving
On the Mac, optional background helpers can continue working while you chat and deliver their reports when done.

Agents settings with a model pool, capability notes and reasoning depth choices (mockup) — Agents & model pools · Interface mockup. Click to explore.

05 / THE WAY IT WORKS

Your habits.
Its instructions.

Turn a useful set of instructions into a reusable skill. Keep your preferred writing style, meeting format or explanation approach close at hand.

Markdown skills@ mentionsCustom personas

Explore the controls

Edit plain Markdown
Add and manage skills in Settings, or edit their Markdown files. Enable only the skills you want JOSIE to use.
Call one in when it matters
Type @S.name to attach a skill to a message. The @. picker helps you find both skills and specialist agents.
Give JOSIE a personality
Choose a persona or write custom instructions, with a voice and orb colour to match. Voice and Chat can have different prompts.

Skills settings with reusable instruction files and individual enable switches (mockup) — Skills & personality · Interface mockup. Click to explore.

06 / BEYOND THE WINDOW

Your Mac.
Beyond the app.

Turn on JOSIE’s optional OpenAI-compatible server to use the Mac’s configured pipeline from other apps, scripts, or the iPhone and iPad preview.

Chat & audioEmbeddingsPhone ↔ Mac

Explore the controls

More than text
Endpoints cover chat completions, speech recognition, speech synthesis, embeddings and realtime sessions.
Choose where it is reachable
The server starts off and binds to localhost when enabled. LAN access and API keys are controlled in Settings.
Bring the Mac to your phone
Add your Mac as a provider on iPhone or iPad over your network or Tailscale. The Mac hosts the models and agent features.

API server settings with LAN access, port and a hidden API key (mockup) — Your Mac as an endpoint · Interface mockup. Click to explore.

THE WORK, IN VIEW

Follow the steps.
Stay in the conversation.

A tool activity pill shows what JOSIE is doing. Expand it to inspect tool results, specialist runs and reasoning when enabled. Set a step limit, answer confirmation cards, or stop a run.

YOUAsk a questionJOSIE + SPECIALISTSSearch · read · checkBACK TO YOUAn answer with the work behind it

A completed agent reply with expanded tool steps and a nested researcher run (mockup) — Open the activity pill to see the work behind the reply. (mockup)

The chat mention picker showing reusable skills and specialist agents (mockup) — Type @. to bring a skill or specialist into the conversation. (mockup)

06 · Screenshots

Thoughtful,inside and out.

The Mac app, plus a preview of the iPhone app. Click any picture to see it larger.

JOSIE chat interface with a conversation and document attachment (mockup)

YOUR SPACE TO THINK

A familiar feeling.
A new kind of company.

Voice, chat, your library, and a voice studio. Everything has its place.

Interface preview · Mockup

Explore every screen

Mac

Calls & chat

Voice mode during a call with transcripts hidden: the big centred orb, a waveform and Listening… under it, and five round call controls with labels (mockup) — Voice: just talk, and JOSIE answers out loud. (mockup)

Voice call with transcripts shown: the small orb at the top, a few turns of transcript with a tool call, and the five controls in a glass capsule (mockup) — Live transcripts appear as you speak. (mockup)

Chat mode with the sessions sidebar, glass bubbles with avatars and timestamps, and a PDF attachment chip in the composer (mockup) — Chat: type when you can't talk. (mockup)

The call window while a PDF is dragged onto it: a dashed drop frame and attachment chips above the controls (mockup) — Drop a PDF onto a call to talk about it. (mockup)

A call with the Camera control on and JOSIE answering about what the camera shows (mockup) — Turn on the camera and ask what it sees. (mockup)

A call after a timed mute: the JOSIE muted chip with a countdown at the top of the call screen (mockup) — “Be quiet for ten minutes”: a timed mute. (mockup)

A call with the developer debug overlay in the top-left corner: call state, end of turn, speech recognition, first token and first audio times, levels and turns (mockup) — Developer overlay: live latency for every turn. (mockup)

Voice mode at golden hour: the warm evening sky behind the orb (mockup) — Golden hour behind the orb. (mockup)

Voice mode at night: a dark sky with stars behind the orb (mockup) — Stars at night. (mockup)

Library & embeddings

The Library window: semantic search, kind, type and date filters, a list of documents and conversations, and one document's details and chunk preview (mockup) — Every document and conversation, searchable in the Library. (mockup)

The Library's Map tab in 2D with PCA: coloured nodes linked to their nearest neighbours and a hover tooltip on one node (mockup) — The embedding map in 2D. (mockup)

The embedding map in 3D with UMAP, orbited so the depth of the clusters shows (mockup) — The same map in 3D. (mockup)

Voice studio & Voice ID

Personalization, Voice: the voice JOSIE speaks as, a style field, a saved voice, and the Clone a voice and Design a voice sections (mockup) — Pick a preset, clone a voice or design a new one. (mockup)

Teach JOSIE your voice with three phrases. (mockup)

Dashboard

Dashboard calls: stat tiles, a call-length histogram, minutes per day and the Who talks chart (mockup) — Calls: lengths, minutes per day and who talks. (mockup)

Dashboard chat: stat tiles, token tiles with total, input, output and min, max and average per reply, and messages per day for you and JOSIE (mockup) — Chat: tokens per reply and messages per day. (mockup)

Dashboard speakers: tiles, the owner's card with similarity bins and match confidence (mockup) — Speakers: how well JOSIE knows each voice. (mockup)

Dashboard latency: per-stage median bars with p90 ticks and recent turns (mockup) — Latency for every stage of a voice turn. (mockup)

Dashboard embeddings: chunk, vector, document and session counts, documents by type and the embedding map panel (mockup) — What's indexed for search. (mockup)

Dashboard export for training: the Record my calls switch, a format picker and the Export button (mockup) — Export your calls for training. (mockup)

Menu bar & API server

The menu bar extra open during a call: status, elapsed time, the last transcript line and menu items for ending the call and muting (mockup) — Calls and mute right from the menu bar. (mockup)

Everything else lives under your avatar. (mockup)

Dashboard API server: status, URL, uptime, sessions, requests per endpoint and recent requests (mockup) — Watch the API server at work. (mockup)

Settings

Settings, Embeddings: the weights picker on 8-bit, the Downloaded badge and the semantic search status line (mockup) — EmbeddingGemma for semantic search. (mockup)

Settings, Conversation: turn taking speed, Smart Turn and speculative switches, language, the barge-in slider and speakers (mockup) — Turn taking, barge-in and speakers. (mockup)

Settings, Tools and MCP: built-in tool switches and Apple apps with permission status and a Grant button (mockup) — Built-in and Apple tools, each with its own switch. (mockup)

Settings, Documents and search: conversation indexing, auto context and the documents list (mockup) — Documents & search. (mockup)

Settings, Retrieval: chunk size, overlap, passages and search limit sliders with Reset to defaults (mockup) — Fine-tune retrieval. (mockup)

Personalization, Assistant: the self-control tools switch with per-tool switches and the While muted section (mockup) — Self-control tools and mutes. (mockup)

Personalization, Appearance: dynamic sky, stars at night, show in menu bar and show transcripts switches (mockup) — Dynamic sky, stars and the menu bar. (mockup)

Settings, Diagnostics: recent tool calls with status dots and durations, log buttons and device info (mockup) — Diagnostics: recent tool calls and logs. (mockup)

Onboarding

iPhone

On iPhone, calls use OpenAI GPT-Live, your own compatible server or JOSIE-Voice on your Mac, and no models run on the phone. It isn't available as a download yet.

GPT-Live calls

The iPhone app during a GPT-Live call: the orb, a short transcript and the call, attach and keyboard buttons (mockup) — A GPT-Live call on the iPhone. (mockup)

An iPhone call through JOSIE-Voice on the Mac: a photo of a plant in the transcript, Josie's answer about it, and the extra Show button in the call bar (mockup) — Calling your Mac: show Josie a photo. (mockup)

The iPhone Lock Screen with JOSIE's Live Activity: state, call timer, last line and an End button (mockup) — The call on the Lock Screen. (mockup)

The expanded Dynamic Island during a call: state, timer, last line and an End button (mockup) — End the call from the Dynamic Island. (mockup)

Onboarding

Settings & dashboard

iPhone Settings, Providers: the built-in OpenAI entry and JOSIE-Voice on my Mac selected for calls, with buttons to add the Mac or another provider (mockup) — Providers: OpenAI, your own server or your Mac. (mockup)

iPhone Dashboard, Calls after a couple of GPT-Live calls (mockup) — Your calls on the iPhone dashboard. (mockup)

Every screenshot and clip on this page is a mockup rendered from the app's source with demo content, not a capture of a running build.

07 · Good to know

Smalldetails.

01

Works offline

After setup, everything works offline except the language model, which runs in Ollama on your Mac.

02

Other models

In Settings, switch to any OpenAI-compatible service with your own key, or to Apple Intelligence on macOS 26.

03

Run setup again

Settings → Setup → Run setup again.

04

Apple tools

Turn them on one by one in Settings → Apple tools. macOS asks permission the first time each is used.

05

Optional extras

They download only if you want them: voice design (3.1 GB) and word timestamps (1.3 GB).

06

For tinkerers

Dashboard → Chat counts tokens per reply. Settings → Developer → Debug overlay shows live latency on the call screen.

08 · FAQ

Questions,answered.

Is it free?

Yes.

Does my voice go to the cloud?

No. Hearing, voice recognition and speech run on your Mac, and by default the language model (Ollama) does too. If you switch to an online model, your text goes to that service.

Do I need an account?

No.

Does it work on Intel Macs?

No, Apple Silicon (M1 or later) only.

Is there an iPhone or iPad version?

Not as a download yet. On iPhone and iPad, calls use OpenAI GPT-Live, or JOSIE-Voice on your Mac over your network or Tailscale. See the iPhone preview (mockups) ↑

It says "Ollama isn't running"

Open Ollama from Applications, then click Recheck in JOSIE's setup.

JOSIE can't hear me

Open System Settings → Privacy & Security → Microphone and turn on JOSIE-Voice.

How much space does it need?

About 5 GB of voice models plus about 1 GB for the language model, at the default small sizes.

How do I uninstall it?

Drag JOSIE-Voice to the Trash. To remove its data and models too, delete the JOSIE-Voice folder in ~/Library/Application Support.

09 · The JOSIE family

JOSIE-Voice comes from the JOSIE family of personality-driven language models by Gökdeniz Gülmez, trained entirely on Apple Silicon.

Meet the JOSIE models ↗ Hugging Face ↗ Sponsor on GitHub ↗

Ready tosay hello?

Download for Mac

The latest DMG from GitHub Releases · Apple Silicon · macOS 14+

Your Mac.With a voice.

A conversation.Not a command line.

Six minutes.One uninterrupted call.

Less to do.More to your day.

Reminders

Calendar

Maps

Messages

Small presence.So many possibilities.

Real conversation

Knows your voice

Voices you choose

Private by design

Takes real action

Sees what you show it

Adapts to the room

Living sky

Personas

Your iPhone, your Mac

Choose how it hears

Its own chat prompt

From your voice.Back to you.

Find the words.

Work through the request.

Keep it flowing.

Make JOSIEyour own.

A mind ofyour choosing.

Pick your backend

Tune how it thinks

Choose how it hears and speaks

Connected.On your terms.

Choose what is available

Set the depth of access

Extend with MCP

Pick upthe thread.

Remember, retrieve, forget

Search your own context

Tune what comes back

A little help.A lot more done.

Make specialists your own

Match the model to the task

Keep the conversation moving

Your habits.Its instructions.

Edit plain Markdown

Call one in when it matters

Give JOSIE a personality

Your Mac.Beyond the app.

More than text

Choose where it is reachable

Bring the Mac to your phone

Follow the steps.Stay in the conversation.

Thoughtful,inside and out.

A familiar feeling.A new kind of company.

Mac

Calls & chat

Library & embeddings

Voice studio & Voice ID

Dashboard

Menu bar & API server

Settings

Onboarding

iPhone

GPT-Live calls

Onboarding

Settings & dashboard

Smalldetails.

Works offline

Other models

Run setup again

Apple tools

Optional extras

For tinkerers

Questions,answered.

Ready tosay hello?

Your Mac.
With a voice.

Six minutes.
One uninterrupted call.

A mind of
your choosing.

Connected.
On your terms.

Pick up
the thread.

A little help.
A lot more done.

Your habits.
Its instructions.

Your Mac.
Beyond the app.

Follow the steps.
Stay in the conversation.

A familiar feeling.
A new kind of company.