Native on Apple Silicon
A companion.
A researcher.
A builder.
Three things people usually rent from a cloud, running natively on your own Mac. Kynd Agent remembers you, digs across the web while you watch, and writes real software — fixing its own mistakes as it goes. Built on MLX, Apple's machine-learning framework, not a wrapper around a server.
Getting there costs about five minutes of your attention.
Apple Silicon (M1 or later) · macOS 13+ · 16 GB memory or more
Setup
Five minutes of your attention.
One download, 919 MB. You drag it to Applications, the way Mac software used to work — Python, the model runtime and the language servers are already inside. No terminal. No pip. No configuration files. On first run, a guided setup looks at your Mac, picks the right model for the memory it finds, and fetches it. There's a progress bar, not a terminal.
A word of honesty: the model itself is several gigabytes, and your connection decides how long that part takes. But it doesn't need you for it. Your share — the clicking, the choosing — is a few minutes. Start the download and put the kettle on.
01 — The companion
Someone to think
out loud with.
You've given Kynd the files you never quite get through — the survey on the house, the pension statement, your own notes. One evening you ask it, out loud, what the survey actually said about the roof. Whisper does the listening, running on your Mac. Kynd checks its library of your files and answers back in a voice. At no point does any of it touch the internet.
It remembers you between sessions, too — what you've told it carries forward, so you're not reintroducing yourself every week. And for conversations you'd rather weren't kept at all, there's a private chats mode.
The built-in macOS voice is free, forever. The licence adds four neural engines, including voice cloning from a short recording — still entirely on-device.
Voices
- Systembuilt into macOS · free forever
- Kokoro0.5 GB · warm and natural, best on 16 GB Macs
- Qwen3 TTS1.7 GB · more expressive range
- Chatterbox3.5 GB · clones from a recording
- Fish Audio S27 GB · the most faithful cloning
02 — The researcher
A researcher you
can watch work.
You need the real picture on something — a grant you might qualify for, a market you're sizing, a decision currently spread across fourteen open tabs. You switch web search on for this one conversation, hand the question to Deep Research, and then you do the good part: you watch.
A live panel shows the agents at work — the searches they run, the pages they visit, the notes stacking up as the dig goes deeper. It is not a black box that hands back a paragraph; it is a researcher you can look over the shoulder of. Like watching a good librarian at speed.
Web search is opt-in, per conversation. Every other conversation stays in the room.
Deep Research — live
searchsmall business energy grants UK 2026
visitgov.uk — business energy advice service
noteeligibility: under 250 staff, meter type…
searchregional variants · Scotland / Wales
visitofgem.gov.uk — non-domestic tariffs
notewriting summary — 2 sources conflict, flagging…
The shape of the live panel: every search, visit and note, as it happens.
03 — The builder
A builder that fixes its own mistakes.
You describe the app you keep meaning to make. Kynd plans the work as a checklist you can read, then gets on with it — writing files, running commands, reading its own error messages and fixing them. Live Preview renders the page while it's still being written, so you watch the thing take shape. And every change it makes has a one-button Undo.
When you'd rather point than describe, WebDesigner lets you edit the page visually. When you'd rather type, there's a real code editor and a shared shell — you and the agent work in the same place instead of passing notes.
A marketplace of agent tool packs extends what it can do, with audio and video generation add-ons to follow.
Quick Start templates
- iOS AppSwiftUI
- iOS GameSpriteKit
- macOS Appnative
- Web AppReact + TypeScript
- Web GameCanvas
- Mobile AppCapacitor · builds iOS and Android
- Landing PageHTML/CSS
Under the bonnet
Tested on your Mac,
by your Mac.
If you've ever stared at a page of quantisation options and sampler settings and quietly closed the tab, this is the part built for you. None of that is intelligence. It is plumbing — and Kynd ships with the plumbing done.
Setup reads your Mac's memory and chooses accordingly. A 16 GB Mac gets a quick 9-billion-parameter model. A 32 GB Mac gets a 35B. A 64 GB Mac gets the same 35B at higher quality. The context window — how much the model can hold in mind at once — scales with memory too, from 16k to 128k.
Then Kynd measures rather than assumes: it benchmarks the actual machine — real generation speed, real memory headroom, real tool calls — and sets its own limits from those measurements. You can race two models head-to-head, graded against fixed expected answers — not vibes.
The defaults are defaults because the alternatives failed. Models squeezed below 4-bit precision pass quizzes but fall apart at real agent work — we measured it, and we don't ship them. We tested the models so you don't have to.
Privacy
Private in ways you can check.
Most privacy pages ask for your trust. This one would rather you checked.
- No account and no sign-in — the app never asks who you are.
- No telemetry on launch — start the app and it calls no one.
- Your chats, projects and memory are plain files in your home folder. Open them, back them up, delete them — they are yours.
- Web search stays off until you switch it on, and then only for that conversation.
- The honest test: pull the network cable out. Kynd keeps working.
Every claim on this list is checkable on your own machine. Not a policy promise — just how it's built.
Price
One payment. Yours for good.
The trial is the full product, free for 30 days. After that, a genuinely useful tier stays free forever — and nothing is ever taken away at the thirty-day mark: your chats, projects and memory are plain files on your Mac and stay exactly where they are. The app simply settles into the free tier until you decide.
| Free | Licence | |
|---|---|---|
| Chat with a local model | yes | yes |
| Web search and web fetch | yes | yes |
| Calibration and benchmarks | yes | yes |
| Knowledge — files it can draw on | up to 10 files | unlimited |
| Memory — what it recalls about you | a rolling 30 days | kept for good |
| Voice | macOS system voice | four neural engines, plus cloning |
| Build — writes files, runs commands | — | yes |
| WebDesigner — edit a page visually | — | yes |
| Deep Research | — | yes |
| Image generation | — | yes |
| Workflows and multi-agent runs | — | yes |
| Workflow and tool builders | — | yes |
$99
Launch price · rising to $149 — so early is the good time
- One payment. Not a subscription. No meter.
- Twelve months of updates included.
- The version you buy is yours and works forever.
- The free tier never expires, whatever you decide.
919 MB · Apple Silicon · the model downloads separately on first run
Questions
Before you download.
Will it run well on a 16 GB MacBook Air?
If you have an Apple Silicon Mac with 16 GB of memory or more, you're in. A 16 GB machine gets the 9B model — tested for exactly this tier — and Calibration tunes the limits to what your specific Mac measures, not what a spec sheet guesses.
Is it as good as Claude or GPT-5?
On raw reasoning, no — a local model is not going to match a frontier model on someone's server rack, and anyone telling you otherwise is selling something. What it is good at is the large middle of daily work: reading your files, researching, making the edits you describe, fixing its own errors. And it does that on a plane, on a train, and on things you are not allowed to paste into a browser.
How big is the download, really?
The app is 919 MB, because the runtime it needs is bundled — that is what makes the no-terminal setup possible. The model is separate and downloads on first run: 7.7 GB for the 16 GB tier, more for the larger models. There's a progress bar, and it doesn't need you to watch it.
Does it work offline?
Yes — that is the point. Once the model is downloaded, chat, voice, memory, and building all run with no connection at all. Web search and Deep Research are the two features that reach the internet, and both are opt-in, per conversation.
What happens when my 30 days are up?
It does not stop. Chat, web search, calibration and the system voice keep working for as long as you have the app, with nothing throttled and no countdown in the corner. What locks is the agent side — Build, WebDesigner, Deep Research, image generation, workflows, the builders and the neural voices — and your knowledge library settles back to ten files while memory returns to a rolling thirty days. Nothing is deleted when that happens: older memories move to the archive on your own disk and come straight back if you ever unlock them.