Yujin

Frontier tech for the next decade of human-computer interaction.

We build software that talks back, listens, and acts on what it hears, instead of asking humans to translate themselves into clicks and typing.

友人

What comes after the form.

Yujin works on the boundary where AI meets the user interface. The interfaces we ship today still assume a human in front of a screen filling fields. We design for a world where agents, voice runners, accessibility tools and humans drive the same surface through the same contract.

船出
Currently shipping

NAC3 -- Native Agent Contract

v2.3.0 · Apache 2.0 · Free

An open standard that lets any web UI be driven by AI agents, voice assistants, screen readers, RPA bots, and Playwright specs -- through the same contract. Five HTML attributes. No SDK lock-in. Free to use, free to fork.

Now in alpha · USD 10 / mo · 7-day free trial

Yujin Forge -- voice-first NAC3 dev framework

v0.2.0-alpha.4 · BYOK · Commercial

The dev tool that turns a Claude or OpenAI or Gemini key into a working NAC3 React app: CLI + browser chat panel + voice loop + spec ingest from ten document formats. Bring your own LLM key, multi-provider routing, end-to-end encrypted key sync between desktop and a mobile companion. $10/month, cancel anytime.

Free · Open source (MIT)

Yuemail -- voice-first email for accessibility

v0.6.0 · MIT · Free & open

A voice-first, single-user, local-only email client for people with motor or visual disability: dictate a document, sign it, and send it by email -- no sighted helper, no precise mouse. It runs on an AI brain, and thanks to NAC3 a cheap economy model is enough -- no frontier model, so it costs cents. Hearing and speaking work in the browser, free.

Free · Open source (MIT)

NAC3 Kikoe -- local voice commands for atypical speech

v0.2.0 · MIT · Zero dependencies

Local, speaker-dependent voice-command recognition for atypical speech (dysarthria, non-standard pronunciation). It recognizes a handful of your own commands by mathematical proximity (MFCC + DTW), fully on-device. The raw audio is discarded after the numeric fingerprint is extracted -- the voice is never stored. Zero dependencies, Node 22+, and 11 native NAC3 verbs. The reference adopter is Yuemail.

Demo script comprando zapatillas via NAC3.

A scripted buyer (pre-recorded TTS) drives Atlas Pro through the same five slot-fills any human visitor would: size, color, qty, delivery, confirm. The seller is Claude / Gemini, the surface is NAC3. Same demo, three languages.

Atlas Pro -- script demo
Video drops here as soon as the recording lands. In the meantime, try the live demo directly.

Try the live demo

AgenteIA comprando zapatillas via NAC3.

No script this time. A second AI (the buyer) talks to the first AI (the seller) through Atlas Pro's voice surface, exactly the way a human would: the buyer listens to each seller turn via STT, replies via TTS, and never reaches into the page's internals -- just the public NAC3 contract. Two agents, one transaction, three languages.

Atlas Pro -- AI buyer agent
Video drops here as soon as the recording lands.

Try the live demo

続く

What's next.

Yujin Pilot -- the voice + chat controller that drives every NAC3-compliant app in the registry -- and a public registry for NAC3 apps. Get notified.