The AI that doesn't just answer. It sees, listens, and acts — right on your machine.
IRIS is a native desktop intelligence built on Rust and React with a governed execution core. Speak a command and watch it ripple across your operating system — apps launch, windows snap, files organize, calendars fill themselves in. And the heart of IRIS is open source.
IRIS is not a chatbot with tools. It is an operating layer for your computer.
What makes IRIS different
Most AI assistants live in a browser tab. IRIS lives in your computer — with permissioned access to your apps, your screen, your voice, and your workflow.
Always-on wake words — "Hey IRIS", "Okay IRIS", "Yo IRIS". Full-duplex voice with native audio capture, speaker verification, and streaming TTS that talks back in real time.
Vision analysis of your webcam and desktop, presence detection that knows when you're at the keyboard, and screen annotations that point at exactly the thing you asked about.
Sixty-plus guardrailed tools spanning app control, web, email, calendar, media, files, mouse, keyboard, and system settings — executed through a risk-assessed pipeline, never raw and unsupervised.
Personal continuity through Lethe and shared memory through Mnemosyne. IRIS builds context about how you work — personal context first, shared knowledge second.
Every action passes a multi-stage safety pipeline: risk scoring, pre-flight simulation, explicit approval for dangerous operations, and full audit logging. Authority is earned, never assumed.
The core of IRIS is open source. Read the code, run it yourself, build on it. github.com/solsticeaistudio/iris-desktop-oss
Capabilities
A sample of what IRIS handles natively. The full capability catalog goes much deeper.
"Open Chrome." "Set brightness to 50%." "Search Gmail for invoice." "Schedule a meeting." Natural language in, real actions out.
"Save this workspace as coding." "Load writing mode." Entire window arrangements restored with one phrase.
"Deploy." Git push, tests, notification — chained YAML macros with 30+ action types, triggered by voice.
Battery warnings before disaster, meeting nudges five minutes out, break reminders when you've been heads-down too long.
Open palm to listen, thumbs up to confirm, shush to interrupt. MediaPipe-powered hands-free control with 20 action types.
Paste a flight number and get status. Paste an API key and get a warning. Fifty-plus entity types recognized in real time.
Gmail and Outlook search plus attachment downloads, calendar queries and event creation, watch rules that alert you proactively.
Live HTML previews, Mermaid diagrams, charts in a side window — and gesture-controlled 3D model viewing in another.
In action
Real demonstrations, zero staging — from wake word to executed command.
IRIS Overview
Full demonstration of IRIS capabilities
IRIS Vision
Screen analysis and visual understanding
Desktop Vision
Visual environment perception
Paper to App
Convert physical documents to digital
IRIS Webpage
Web interaction demo
Calendar Scheduling
Voice-controlled calendar management
Spotify Launch
Voice app launching
YouTube Launch
Voice web navigation
Brightness Control
Voice display adjustment
Volume Control
Voice audio adjustment
IRIS Move
Window positioning commands
Architecture
IRIS isn't a prompt wrapped around a script. It's a three-layer native stack topped by a deterministic governance layer that treats AI authority as a first-class security problem.
Open source
We believe a desktop agent you can trust starts with code you can read. That's why IRIS's core lives in the open — the same ideas, the same guardrail philosophy, free for developers, researchers, and builders everywhere.
Risk-tiered execution, approval gates, and audit logging — inspect exactly how IRIS decides what it's allowed to do.
Fork the patterns: tool registries, macro engines, memory wiring, voice pipelines. Stand on the core instead of rebuilding it.
An agent with this much reach owes you transparency. Open source isn't a feature of IRIS — it's the accountability model.
Trust by design
An agent that can touch everything should be able to do nothing quietly. Every IRIS action walks the same gauntlet:
Every tool carries a danger level from low to critical — behavior adapts accordingly.
Dangerous actions are simulated first: affected files, data at risk, reversibility.
High and critical operations wait for a human yes. Denial cancels — no retries.
Intent, approval, result — logged durably so every action can be explained later.