What Claude Computer Control Actually Is

Anthropic has a feature called Computer Control (sometimes called Computer Use) that lets Claude interact with a computer the way a human does — moving the mouse, typing, clicking buttons, reading screens.

It's one of the most consequential capabilities they've built. And understanding how it works technically matters if you're designing agentic workflows.

What It Actually Does

When you give Claude access to computer control tools, the model can:

• → Move the cursor and click on screen elements
• → Type text into fields and applications
• → Read screenshots and understand what's on screen
• → Execute shell commands and run applications
• → Scroll, drag, and interact with any UI element

Instead of just generating text or calling APIs, Claude can now work inside a graphical user interface. It sees what a human sees and acts on it.

This is different from traditional tool use. With code execution tools, the model runs a bash command and gets stdout back. With computer control, the model sees pixels and decides what to click.

How It Works Technically

Computer control is built on top of Anthropic's Tool Use API. The model generates tool calls that represent computer actions, and a runtime executes them against a screen environment.

The tool loop looks like this:

1. Claude observes screen state (via screenshot)
2. It decides on an action (click here, type this, scroll there)
3. It outputs a tool call with the action parameters
4. The runtime executes the action against the real screen
5. A new screenshot is captured
6. Loop back to step 1

The model isn't just following scripts. It's reasoning about UI state the same way a human would — "that button looks disabled, let me scroll down first" or "I need to click the text field before typing."

Anthropic acquired a company called Vercept specifically to advance these capabilities, which tells you how technically complex screen interaction is compared to API calls.

Which Models Support It

Claude Opus 4.6 is the flagship model for computer use. From Anthropic's own reporting:

> "Across agentic coding, computer use, tool use, search, and finance, Opus 4.6 is an industry-leading model."

The full 4.6 family (Opus, Sonnet, Haiku) supports varying levels of agentic capability, but Opus is the one designed for full computer control tasks where reliability matters.

If you're using Claude Code or the Agent SDK with Opus 4.6, you have access to computer control. It's not a separate product or add-on — it's built into the model's tool use capability when you give it the right permissions.

Can You Use It via Minimax or Ollama?

Short answer: No.

Minimax is a separate AI provider with its own API schema. Computer Control is tied to Anthropic's specific tool definitions and model capabilities. Minimax doesn't expose Claude Opus 4.6 or Anthropic's tool schema.

Ollama runs open-source models locally (Llama, Mistral, Qwen, etc.). Those models don't have Claude's computer use capability, and Ollama's runtime doesn't support Anthropic's tool protocol.

What does work beyond Anthropic directly:

• → Amazon Bedrock — Claude via AWS infrastructure
• → Google Vertex AI — Claude via Google Cloud
• → Microsoft Azure AI Foundry — Claude via Azure

These are official partnerships where Anthropic's models and tool protocols run on cloud infrastructure you already use. If you're already in AWS or GCP, this is the path.

Why This Matters for Agent Workflows

Computer control is the final piece that makes agents truly autonomous. You can have an agent that writes code, runs tests, and deploys via APIs — but if something requires clicking through a GUI or interacting with a legacy system, you're stuck.

With computer control:

• → Agents can automate anything a human can do on a screen
• → Legacy applications with no API become automatable
• → Testing that requires browser interaction works natively
• → Desktop applications are fair game

The catch — and it's a real one — is that computer control at full autonomy is risky. An agent clicking through your cloud console or moving files can cause real damage. This is exactly why human-in-the-loop oversight matters so much at this level of capability.

Where AgentRQ Fits In

AgentRQ is built for this moment. When Claude's computer control actions hit a decision point — a confirmation dialog, a risky operation, a judgment call — AgentRQ surfaces that decision to you in real time.

The agent doesn't stop. It pauses the specific action, sends you a notification with context, and waits for your signal. You approve or redirect from anywhere, and the agent continues.

This is Level 5 autonomy: the agent works on your behalf continuously, with your judgment applied exactly where it matters and nowhere else.

Connect Claude Code to AgentRQ and set up human oversight for your computer control workflows.