Four ways to run it
One runtime, four front doors. Pick the one that matches where the task comes from.
Inference is your call too: H’s hosted Models API by default, or any OpenAI-compatible server on your own hardware with
--base-url, in which case no screenshot, keystroke, or app content leaves your machine.
Start here
Install, pick how you launch it, then tune it to your setup.Supported platforms
Theholo client is pure Python and runs anywhere Python does. The hai-agent-runtime binary it drives is downloaded on first run where H publishes it, and bring-your-own elsewhere.
macOS, Apple Silicon
Runtime downloaded on first run. Needs Screen Recording and Accessibility permissions.
Windows, x86_64 and ARM64
Runtime downloaded on first run. No extra permissions.
Linux, x86_64
Runtime downloaded on first run. Needs an X11 session. Wayland is not supported.
macOS, Intel
Bring your own runtime. Put
hai-agent-runtime on PATH, or set a download URL.holo client normally, then put a hai-agent-runtime executable on PATH, or point HAI_AGENT_RUNTIME_DOWNLOAD_URL and HAI_AGENT_RUNTIME_DOWNLOAD_SHA256 at a trusted build. The CLI checks PATH first, so your binary wins over the managed one. See Runtime cache.
What’s next
Two things are coming, with no firm dates yet:- Background mode. Today the agent drives the screen in front of you and shares your mouse and keyboard. Next, it binds to its own window so you keep using your machine while it works.
- Native app. The same agent as a desktop app beside your daily workflow, no terminal needed.