Iris

What Iris does

Iris uses a two-brain design. Gemini Live handles realtime voice, including interruptions, while Hermes Agent does tool-heavy work such as research, coding, files, terminal and browser automation. Tasks are delegated in the background and return a run_id immediately, so the conversation never blocks, and each Hermes tool call streams into a task card with durations. When a task finishes, Iris announces and summarizes the result aloud.

The Hermes interaction bridge surfaces clarification choices, dangerous-command approvals, sudo passwords and secret prompts in Iris, and sensitive values stay in the UI and never enter Gemini. A confirmation gate reads back a proposal before dispatch. Other features include a local Hey Iris wake word using on-device ONNX inference, MediaPipe hand tracking for clicking, scrolling and closing in the air, and a Glass HUD mode that floats a transparent click-through overlay above the desktop. A bounded snapshot of Hermes USER.md and MEMORY.md keeps conversation coherent.

Key features

• Gemini Live realtime voice with barge-in and WebRTC echo cancellation
• Background delegation to Hermes with a run_id and a live feed of each tool call
• Interaction bridge for clarifications, dangerous-command approvals, sudo passwords and secret prompts
• Local Hey Iris wake word using on-device ONNX inference
• MediaPipe hand tracking to click, scroll and close the UI in the air
• Glass HUD mode, a transparent click-through overlay above the desktop

When to use it

• Talking to Hermes hands-free while it works on research or coding tasks in the background
• Approving dangerous commands or answering clarification prompts from the Iris interface
• Keeping a voice companion floating over your desktop while you keep working

Who it is for: Hermes Agent users who want a voice-first desktop front end for delegating work.

How it fits with Hermes Agent

Iris uses Hermes Agent as its work engine, delegating tasks to it and reading a snapshot of its USER.md and MEMORY.md for shared context.

FAQ