DocumentationTry Hanzo
Computer use, in a container

Hanzo Operative

A container with a Linux desktop inside it, and a model driving that desktop by screenshot, mouse and keyboard. You watch through a browser and send instructions from the same page. Nothing it does reaches your machine, because the machine it is using is the container.

Screen Vision
Multimodal Models
Cursor Control
Keyboard Actions

What is in the container

A screen, a way to watch it, and the tools the model drives it with

A desktop in the image

Xvfb and X11 give the container a screen with no monitor attached, so a real window manager and real applications run inside it.

Watch it work

VNC on 5900 for a native client, noVNC on 6080 in a browser, and the control page on 8501. You see the cursor move as it moves.

Say what you want

Type an objective and it plans, acts, looks at the result, and goes again. You can interrupt at any point.

Bash and a file editor too

Not only the screen. It can run a command or edit a file directly when that is the shorter path, instead of typing into a GUI text editor.

Blast radius is the container

Stop it and everything it touched goes with it. Give it the narrowest privileges and an allowlist of domains, and keep credentials that matter out of the image.

Where the model comes from

The API directly, AWS Bedrock, or Google Vertex — chosen with an environment variable. The container is the same in all three cases.

Open Source onGitHub

One docker run

Nothing is installed on your machine. The desktop, the tools and the agent are all inside the image

Installation

Pull the image

docker pull ghcr.io/hanzoai/operative:latest

Run it with your key and the four ports

docker run -e ANTHROPIC_API_KEY -p 8501:8501 -p 6080:6080 -p 5900:5900 -p 8080:8080 -it ghcr.io/hanzoai/operative:latest

Open localhost:8501 and type an objective

System Requirements

  • • Docker. Nothing else installed on the host
  • • A key for the API, Bedrock, or Vertex
  • • Four ports free: 8501, 6080, 5900, 8080
  • • Give the container the fewest privileges that work

Basic Usage

Watch it in a browser, no VNC client needed

open http://localhost:6080

Send the model through Bedrock instead

docker run -e API_PROVIDER=bedrock -e AWS_PROFILE -e AWS_REGION=us-west-2 …

Or through Vertex

docker run -e API_PROVIDER=vertex -e VERTEX_REGION -e VERTEX_PROJECT_ID …

Keep the session between runs

docker run -v $HOME/.anthropic:/home/operative/.anthropic …

Computer use is beta, and instructions written into a web page or an image can try to talk to the model as if they were you. Limit outbound traffic to domains you name, and ask for a human yes before anything with consequences outside the container.

Where the model comes from

One environment variable decides. The container, the desktop and the tools are identical in all four cases

Recommended

Hanzo

by The default

  • Runs through api.hanzo.ai
  • Every run shows up in the console with its cost
  • Bring a Hanzo key, or one you already have
  • Pick the model in the sidebar

docker run -e HANZO_API_KEY …

Direct

by Straight to the model vendor

  • No Hanzo in the request path
  • Your existing key and your existing bill
  • Nothing here to log the run for you
  • Set API_PROVIDER=anthropic

docker run -e API_PROVIDER=anthropic …

Bedrock

by Inside your AWS account

  • Reads your local AWS profile
  • Stays in the region you name
  • For when the model must not leave your account
  • Set API_PROVIDER=bedrock

docker run -e API_PROVIDER=bedrock -e AWS_PROFILE …

Vertex

by Inside your Google Cloud project

  • Application default credentials
  • Names a region and a project id
  • Same container, same desktop, same tools
  • Set API_PROVIDER=vertex

docker run -e API_PROVIDER=vertex -e VERTEX_PROJECT_ID …

Which model a provider offers changes; the container does not. Thedocumentation lists what each one currently accepts.

See Operative in Action

Watch as Hanzo Operative uses multimodal AI to router interfaces, complete tasks, and solve problems autonomously.

Demo Video

Open Source Revenue Sharing

Up to 5% of compute goes back to open source

Every deployment is SBOM-verified. Contributors to Anthropic Computer Use earn a share of compute revenue — transparent, on-chain, and customizable by the community.

Run it and watch

The most useful hour you can spend with computer use is the first one, watching where it gets stuck.

Read before you run it

Open source

License: Apache-2.0hanzoai

Get Operative

Computer-use harness