Kynd Agent
macOS · Apple Silicon
Kynd Agent loads an open model onto your own machine and puts it to work: reading your project, writing files, running commands, and showing you the result as it builds. No cloud round-trip. No per-token meter running in the background.
Apple Silicon (M1 or later) · macOS 13+ · 16 GB memory minimum, 32 GB comfortable
Your code stays on the disk it started on.
Everything below — the reading, the reasoning, the writing — happens on the Mac in front of you. The only time anything leaves is when you explicitly switch on web search, and then it is a search query, not your source.
How it works
Give it a job. It plans the work as a checklist, then goes through it — searching the codebase, editing files, running commands, reading the errors it causes and fixing them. Every file change is undoable from a single button, so a bad run costs you a click rather than an afternoon.
The same model without the file access, for when you want to think out loud, read something back, or ask what a chunk of code actually does. It remembers what you have told it between sessions, and it can speak and listen entirely on-device if you want it to.
The workbench
The right-hand panel is a live view of whatever the agent is doing. HTML it writes renders immediately in Preview. There is a real editor for when you want to take the keyboard back, a visual editor for nudging layout, and a shell you share with the agent.
The screenshot up top is a genuine session: the agent researched open-source animation libraries, built an Electron character animator across eight steps, and the panel on the right is that app running.
Panels
Calibration
Most local-AI tools hand you a settings page and hope you know what a context window is. Kynd Agent runs a benchmark on your actual machine — real speed, real memory, real tool calls — and sets the agent's limits from what it measured.
You can also put two models head to head on identical work and see which one earns its disk space. Every answer is graded against a fixed expected result, so no model is quietly marking its own homework.
Privacy
Price
$49
Launch price · $79 after the first hundred
919 MB · Apple Silicon · the model downloads separately on first run
Before you download
On raw reasoning, no — a 9B model on your laptop 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 a codebase, making edits you describe, wiring things up, fixing its own errors. And it does that on a plane, on a train, and on code you are not allowed to paste into a browser.
That depends entirely on your Mac and the model you pick, which is exactly why Calibration exists rather than a number on a marketing page. For a rough shape: a 9B model on Apple Silicon lands in the low tens of tokens per second — quick enough to read along with, slower than the cloud.
The app bundles the runtime it needs so that the first launch works without you installing Python, a compiler, or a model server. The model itself is separate and downloads on first run, so you pick the size that suits your machine.
No. It leans on Apple Silicon's unified memory to hold the model, and there is no honest way to fake that on an Intel machine.
It stops, and asks you to buy a licence. Nothing is deleted, nothing is held hostage — your projects and chats are plain files and stay where they are whatever you decide.