The first agent-native operating system — where the AI is part of the system, not an app on top of it.
Claw OS is a complete Linux-based environment where the AI agent is a system-level layer, not an application running on top of one. The agent runs as a privileged system daemon (clawd) with direct, scoped access to the kernel, processes, logs, network, and every installed app — so it can reason about and act on the whole machine, the way an operator would, instead of being trapped inside a single app's sandbox.
Agent-native is the whole bet: the next operating system is one where the assistant is not a tab you open but the layer the system runs through. Here that means:
- The AI is part of the system, not an app on top of it. It runs as a privileged system service, not in a sandbox, so you can just ask the things an app could never reach — "why is my network so slow?", "why did that app just crash?", "why can't I get online?", "what's eating my disk?" — and it actually looks at your processes, logs, network, and resources to answer.
- One agent across every app. Instead of a separate, siloed assistant inside each app, a single agent spans the whole machine — it can drive several apps in one request (read a document, draft an email, post a notification) and reasons from real system state, not just one app's view.
- It remembers. The agent has persistent memory across conversations: what you did in your apps, what changed on the system, your preferences, and context that carries from one app to the next — so you never re-explain yourself. It's your memory; you can review it and forget any of it, down to a single app.
- Local-first. The system's intelligence belongs on your own machine, not someone else's servers. As edge hardware (CPU / GPU / NPU) matures, fully on-device "edge AI" becomes the default; today it runs local models where your hardware can handle it and falls back to cloud AI APIs where it can't — without changing how you use it.
- Apps open up to the agent. Every app publishes what it can do, and the agent decides which to use. Developers expose their app to the system agent through the
claw-os-sdk(Rust, Python, Node, Go). - Safe by construction. Every privileged action goes through a capability check and your approval, so "system-level" never means "unrestricted" — looking, changing, and touching the kernel are each gated separately. Primitives return structured, machine-readable data instead of scraped UI, and setting up, asking, chatting, and diagnosing are first-class parts of the OS.
What runs today is the foundation for that future: the system-level agent, its persistent memory, the local-first runtime, the scoped permission model, the cross-app session store, and the app SDK are already here.
Pick an entry point:
- WSL — recommended
- Docker / OrbStack — recommended
- Desktop / ISO / VM — experimental
Import the latest WSL rootfs:
$arch = if ($env:PROCESSOR_ARCHITECTURE -eq "ARM64") { "arm64" } else { "amd64" }
$tarball = "claw-os-wsl-$arch.tar.gz"
$url = "https://github.com/xiaoyu-work/claw-os/releases/download/wsl-latest/$tarball"
Invoke-WebRequest $url -OutFile $tarball
wsl --import claw-os C:\WSL\claw-os .\$tarball --version 2
wsl -d claw-osRun the container:
docker pull ghcr.io/xiaoyu-work/claw-os:latest
docker run -d --name claw --privileged -v ./workspace:/home/cos/workspace ghcr.io/xiaoyu-work/claw-os
docker exec -it --user cos claw bash --loginDesktop images are experimental.
Build a bootable desktop VM disk image and load it in VMware — see Building the Claw OS Desktop Image for the steps on Windows (WSL2), macOS, and Linux.
cos # list primitives
cos app web read https://example.com # fetch URL → {url, title, text, links}cos agent setup # configure providers and credentials
cos agent ask "why is my network so slow right now?"
cos agent ask "why did my last app crash?"
cos agent chat # interactive REPL with cross-app memoryMIT for the kernel and apps. See the NOTICE for vendored sources and their upstream licenses.