A desktop voice agent for Google Workspace and grounded local tasks, combining hosted Gemini Live conversation with local Gemini CLI execution on the connected machine.
Google Workspace workflows with grounded local execution
Relay keeps the conversation live and voice-driven, while making sure execution stays connected to the user's actual desktop state instead of drifting into detached assistant output.
Relay is built for the gap between talking to an assistant and actually getting work done across Google Workspace plus the local desktop. Instead of forcing users to bounce between a chat box, files, browser tabs, and Google tools, Relay keeps those actions inside one voice-first workflow.
Relay exists to test a stronger product pattern than generic chat assistants: keep the live conversation in a hosted core, but let grounded execution happen on the user’s real machine when the task requires local state, files, apps, or browser context.
Relay has a public landing page, a public repository, a public demo video, and downloadable desktop builds. The current public positioning is judges-first and early-access oriented, but the product is already presented as a live desktop agent with a clear setup path, system boundary, and real workflow promise.
The public materials already do useful trust work: the landing page explains the architecture and setup flow, the GitHub repository exposes the implementation scope, and the demo video shows the intended voice-driven workflow before someone installs anything.
These captures show both sides of the public Relay story: the setup and positioning on the landing page, and the live desktop interface shown in the public demo.
The public landing page does more than market the product. It explains the desktop voice-agent promise, the judge-access flow, the quick-start commands, and the download path before installation.
This public demo frame shows the Relay desktop surface itself: the live voice session, the active task runner, and the results panel that keeps background execution visible instead of hiding it behind a chat transcript.
Relay is strongest when the system boundary stays explicit: voice at the desktop surface, orchestration in the hosted core, and grounded actions on the connected machine.
Relay uses an Electron desktop surface for voice input, playback, transcript visibility, and task continuity so the assistant feels conversational instead of form-driven.
The cloud core owns the Gemini Live session, orchestration, and canonical task state instead of hiding product logic inside the local shell.
When a task needs real device context, Relay delegates local execution through Gemini CLI so files, apps, and browser state stay grounded in the actual machine.
With the authorized Workspace extension in place, Relay can connect Docs, Drive, Gmail, and local desktop work without making the user manually stitch those steps together.
Relay is publicly presented through its landing page, repository, demo video, and downloadable builds for macOS and Windows. The current setup still expects Gemini CLI sign-in and Workspace extension authorization before the full workflow is available, and the public landing page currently frames access as judges-first or early access.