1:1 mentoring with Big Tech AI engineers
LLM & Agentic

Computer & Browser Use

Let Claude drive a browser or desktop safely: the computer and browser use toolsets, batched actions run in order, screenshot token costs, element references, and the sandbox, allowlist and human gate that bound a fooled agent.

Last updated

Production10 min readFirst readThe Agent LoopTool Use with Claude

After this section you can

  • Choose between web fetch, the browser toolset and the computer toolset for a task
  • Run a batch of computer actions in order, stop at the first failure and answer every block
  • Size screenshots to the image limit and bound a fooled agent with a sandbox, an allowlist and a human gate
34

Computer & Browser Use: Agents That See and Click

Claude can drive a desktop or a browser, but it never touches either. It proposes clicks and keystrokes as tool calls; your executor runs them in a sandbox and sends back what the screen looks like now.

Key idea

Computer use is the agent loop with eyes. Each turn, Claude returns a batch of actions; your code runs them in order, stops at the first failure, and returns a screenshot or a page tree. Everything that makes it safe lives in that executor and the sandbox around it, because the pages it reads are written by strangers.

Claude proposes actions; your executor runs them and reports what it sees
Claude sees the screen proposes actions Your executor runs calls in order stops at the first failure asks a human first when an action has consequences Sandbox low-privilege VM fresh browser profile domain allowlist 1 · tool_use batch click, type, screenshot 2 · act 3 · observation screenshot or page tree 4 · a tool_result per call each echoes toolset_name Repeat until Claude ends its turn without a tool call.

Related

More in LLM & Agentic

Get full access to all 149 sections with code examples, diagrams, and interactive animations.

Unlock Premium