The voice that uses your computer for you
A voice agent for Windows, heard by AssemblyAI. Made for people who can't see the screen, and for anyone it confuses.
0 false yeses in 60 spoken answers. It never acts on a maybe.
Computers were built for people who can see them 
. A screen reader reads what apps label, and when they don't, all you hear is “button, button, button.” Iris listens instead 
. Ask “did my grandson write?” 
and it reads you the letter. Say “order more vitamins” 
and it fills the basket, names the price, and waits. Nothing that sends, pays or deletes happens until you say yes 
.
The green line is Iris asking first. A “no” ends the task: it doesn't look for another way.
The fast part does almost all the work; the slow part steps in only when it has to. Iris draws its work on the screen, so a sighted helper can follow along.
Streaming speech-to-text. Short pauses merge into one turn, because people, older people especially, pause mid-sentence.
Turns what you said into small steps and picks the tool: an API when one exists, the screen when it doesn't. Out of the box it's Qwen on the AssemblyAI LLM Gateway; bring Claude if you have it. Results are read from the screen, never guessed.
Jev by TypeSafe picks each button from the accessibility tree; with an NVIDIA GPU, its open-weights sibling Kev does it locally in about 230 ms. It opens apps by name and uses the keyboard: Enter sends, Escape closes. With Claude, it also looks at the screen to find buttons nobody labelled.
Anything that sends, pays, deletes or shares waits for a yes, judged on your exact words. “Uh… who is it for?” gets the question again.
The brain is one line in .env: IRIS_BRAIN=cloud (the default, no keys), claude-code (your Claude Code login), anthropic (your API key; Claude also becomes the eyes) or gateway (your AssemblyAI key, any Gateway model). Without Claude, Iris has no vision: it works from the accessibility tree and the keyboard.
Our end-to-end runs show the whole loop works on real Gmail, Telegram and Amazon, so you can try it on your own PC. They are not a promise of production reliability: it depends mostly on the brain. With Claude, 5 of 5 Amazon runs ended safely; a small 4B model on the Gateway answers 3× faster but got 1 of 13 right. Details in bench/results/.
In 60 spoken answers it heard 0 false yeses. In 40 actions it had never seen, it let 0 risky ones through without asking. With Claude as its eyes, it clicked 10 of 10 unlabelled icons; with Kev on your own GPU, it decides in 230 ms. Passwords never reach a model: code types them from Windows Credential Manager. Apps you list, it never touches. Iris Cloud counts, it doesn't listen: your speech goes straight to AssemblyAI, and the server never stores what you say or what's on your screen.
Measured with the benchmarks in bench/ on our own test sets. Numbers from September 2026.
Iris is a Windows program: it works on your screen, in your apps. One line installs it, in about 2 minutes. No keys, no GPU, no git.
You need Windows 10 or 11, a microphone and headphones. That's all: speech, brain and decisions run through Iris Cloud, our small server that keeps the keys.
What the server sees Each PC gets an anonymous token and a daily allowance. The server stores counters only, never your speech or your screen.
irm https://raw.githubusercontent.com/Whom-m0rty/Iris-voice-agent/main/bootstrap.ps1 | iexWant more? Set IRIS_BRAIN=claude-code, anthropic or gateway in .env (or install with install.ps1 -Brain claude-code) to use your own Claude or AssemblyAI key; install.ps1 -Local runs Kev on your own NVIDIA GPU. Details in the README.
