hai-agent-runtime binary locally, connects to it over loopback, and lets you operate the visible desktop from a terminal, an MCP or ACP host, or Python code.
HoloDesktop CLI takes control of the visible desktop for the duration of a task. It can open apps, switch focus, click, type, and observe whatever is on screen until the task completes, times out, or is stopped.To stop an active run, press
Esc twice quickly, press Ctrl+C in the terminal or agent host that launched it, abort the host request, or run uv run holo stop from another terminal. If the normal stop path does not respond, use uv run holo stop --force.Choose your path
Install HoloDesktop CLI
Clone the repo, sync dependencies, and verify the CLI before you run an agent task.
Choose a hosted or local model
Use an H Company hosted model, or point the CLI at a local OpenAI-compatible model server.
Run your first task
Run one safe desktop task, grant permissions, and confirm the CLI can act on your machine.
Use the checklist
Check the model backend, runtime, login, permissions, and logs if the first run fails.
What HoloDesktop CLI can do
Use HoloDesktop CLI when an agent needs to see and operate real desktop applications:- run a one-shot task from your terminal with
holo run; - delegate screen-reading and clicking from an MCP host;
- run as an ACP sub-agent for compatible agent hosts;
- embed sessions from Python with
AgentApiClient.
Before you start
You need:- Git and Python with
uv; - either an H Company account for hosted mode, or a local OpenAI-compatible model server;
- on macOS, Screen Recording and Accessibility permissions for the process that launches the CLI.
holo-desktop from a source checkout. The agent runtime downloads on first run unless it is already on PATH.