Agents that take actions, running entirely on local hardware
Desktop agents that run shell commands, manage files and drive coding sessions, with speech handled on-device. Built to work out where an agent should be allowed to act.
An assistant that answers questions is a different piece of engineering from an agent that does something. The second one can be wrong in ways that cost you.
Three desktop agents, all Tauri shells over Python backends on a WebSocket transport. One is voice-driven: speech-to-text on the GPU, speaker identification, then shell execution, file management, web search and media control. No audio reaches a cloud service at any point. The second wraps coding sessions, with project state, history and version control awareness. The third indexes a document vault and answers across it.
What the work actually produced was not the agents. It was the answers to the questions you only hit once an agent can act: which operations need confirmation and which don't, what a session should remember between runs, how an agent recovers when a tool fails halfway, and how to keep a model that is holding shell access from doing something irreversible.
Those answers were what went into client agent work.
Working on something in this territory? Tell us what you’re building