Namzu × Computer use. Sandboxed desktop control.
@namzu/computer-use exposes desktop control — screenshots, mouse, keyboard, file system — as a Namzu tool. The agent gets a virtual display; the kernel runs the sandbox; the host machine is never touched directly.
$ pnpm add @namzu/sdk @namzu/computer-useComputer use is one of the few agent capabilities where sandboxing is non-negotiable. Vendor SDKs ship the model API; the runtime layer is left as an exercise. Namzu is that runtime layer — process isolation, signal handling, checkpoint/resume across long sessions all come from the kernel rather than from your application code.
- →Sandboxed display via Xvfb or a container, never the host screen
- →Pluggable backends — local sandbox, Docker, Browserbase, E2B
- →Compatible with Claude, GPT, and any vision-capable model
- →Kernel-level signal handling for clean session abort
import { createKernel } from '@namzu/sdk'
import { anthropic } from '@namzu/anthropic'
import { computerUse } from '@namzu/computer-use'
const kernel = createKernel({
provider: anthropic({ model: 'claude-sonnet-4-6' }),
})
const agent = kernel.spawn({
tools: [computerUse({ display: ':1', sandbox: 'process' })],
})
await agent.run('Open the staging site and take a screenshot of the header.')Ship Computer use agents on a runtime you can own.
Computer use owns the model. Namzu owns the runtime. The provider package is a thin adapter, not a wrapper, so you keep everything Computer use ships and add the kernel concerns you do not get from the model API alone.