Open source • Claude computer use on AWS Bedrock

An AI agent that uses a computer like you do.

Describe a task in plain English. Vessel's agent looks at a live Linux desktop, clicks, types, and runs terminal commands inside an isolated sandbox, while you watch it happen in your browser.

Built withNext.jsClaude Sonnet 4.5AWS BedrockVercel SandboxnoVNC

How it works

1. Describe the task

Type what you want done, e.g. “Open Chrome, find top TypeScript repos on GitHub, save a summary.”

2. The agent sees and acts

Vessel screenshots the desktop and sends it to Claude. Claude replies with an action (click, type, or bash command), which runs in the sandbox. This repeats until the task is done.

3. You watch it live

The desktop streams to your browser through noVNC, alongside a log of every tool call the agent made.

Example session: watch the agent work

A sample run showing the tool calls Vessel makes. Launch the dashboard to run a real one.

example-session
Claude Sonnet 4.5
Sample instructions
USER PROMPTEXAMPLE

"Open Chrome, go to github.com/trending/typescript, and tell me the name and description of the top repository."

// Tool calls the agent makes
computerExecute({ action: 'screenshot' })
computerExecute({ action: 'left_click', coordinate: [420, 60] })
computerExecute({ action: 'type', text: 'github.com/trending/typescript' })
computerExecute({ action: 'screenshot' })

Agent reads the page and replies with the top repository's name and description.

Screen: Chrome • GitHub Trending
In the real dashboard, the live desktop appears here via noVNC
POST /api/chat • Single-request agent loopCredentials stay server-side

A real run

Recorded from the live dashboard: one browser task and one terminal task.

Prompts: “Find the top trending TypeScript repo” and “Inspect the system in a terminal”.

Who it's for

Developers building agents

A working reference for Claude computer use: the tool loop, screenshot handling, and message formatting, all server-side.

Teams on AWS Bedrock

See how to keep Anthropic's computer-use beta enabled on Bedrock using the native SDK.

Anyone evaluating agents

Watch what an agent actually does on a real desktop, in a sandbox where mistakes can't touch your machine.

Under the hood

The three engineering decisions that make Vessel work: an isolated desktop, a Bedrock integration that keeps computer use enabled, and a server-side agent loop.

Isolated Desktop Sandboxes

Every session provisions its own Vercel Sandbox running a Linux Xfce desktop and Google Chrome. The agent can't touch your machine, and you watch every action live through a noVNC stream.

Native Bedrock Integration

Calls Bedrock through @anthropic-ai/bedrock-sdk (InvokeModel), so the computer-use-2025-01-24 beta flag reaches the model. Through the Vercel AI SDK's Bedrock provider it didn't, and Claude described actions in text instead of performing them.

Server-Side Agent Loop

The whole tool loop runs inside one POST /api/chat request. Screenshots stay in server memory between steps, and AWS credentials and sandbox tokens never reach the browser.

Get started

Try it in the browser, or run your own copy.

Use the hosted demo

Open the dashboard, pick a sample instruction or write your own, and watch the agent work.

Launch Vessel

Run it yourself

git clone https://github.com/Bishoy-Morgan/ai-agent-dashboard
cd ai-agent-dashboard
pnpm install
# add your AWS + sandbox credentials (see README)
pnpm dev

Frequently Asked Questions

What is Vessel?

Vessel is an open-source dashboard where Claude controls a sandboxed Linux desktop from plain-English instructions. You watch the desktop and the agent's tool calls live in your browser.

How does Vessel control the computer?

It screenshots the virtual desktop, sends the image to Claude Sonnet 4.5, and receives tool_use instructions back. Those run as computerExecute (mouse and keyboard) or bashExecute (terminal) inside an isolated Vercel Sandbox, and the loop repeats until the task is done.

Why use @anthropic-ai/bedrock-sdk instead of the Vercel AI SDK's Bedrock provider?

Anthropic's computer-use beta flag didn't reach the model through the AI SDK's Bedrock path, so Claude narrated actions instead of emitting tool calls. The native SDK sends it correctly.

Do API keys ever reach my browser?

No. The whole agent loop runs server-side in POST /api/chat. AWS credentials and sandbox tokens stay on the server.

Can I run it myself?

Yes. Clone the GitHub repo and follow the README. You'll need your own AWS Bedrock access and a Vercel Sandbox.