Services or capabilities described in AWS documentation might vary by Region. To see the differences applicable to the AWS European Sovereign Cloud Region, see the AWS European Sovereign Cloud User Guide.

Use computer use tools to automate GUI tasks with Amazon Bedrock models

Computer use lets Anthropic Claude models automate graphical user interface (GUI) tasks. The model returns tool_use actions such as taking a screenshot, clicking, typing, and scrolling. Your application executes those actions on a desktop or container and returns the resulting screenshot to Claude on the next turn.

Amazon Bedrock offers computer use through two interfaces:

For the underlying tool protocol, the full list of actions, and prompt-engineering guidance, see Computer use in the Anthropic documentation. This page covers the Amazon Bedrock-specific usage on top of that protocol.

Use computer use on Amazon Bedrock

Computer use is supported on both the bedrock-runtime and bedrock-mantle endpoints. On bedrock-runtime, you can use it with the InvokeModel and Converse APIs. You can enable computer use in one of two ways, depending on the model:

Where you put the beta value depends on the endpoint and API:

The bash (bash_20250124), text editor (text_editor_20250728), and memory (memory_20250818) tools don't require a beta value. For all Anthropic-defined tools and their fields, see Computer use.

Supported models and tool types

The following table shows the computer use tool types that each Claude model accepts on the bedrock-runtime endpoint. Submitting a tool type that a model doesn't accept returns a 400 invalid_request_error. The computer_20241022 tool type isn't accepted by any model in this table.

To find which endpoints each model is available on, see the model's Models at a glance.

Example requests

Each of the following examples asks the model to take a screenshot. The model responds with stop_reason set to tool_use and a tool_use block that describes the action. Your application runs the action and returns the result in a tool_result block on the next turn.

InvokeModel with the computer tool

The following Python example sends a request through the bedrock-runtime endpoint with the computer_20251124 tool and the required beta value.

import boto3
import json

client = boto3.client("bedrock-runtime", region_name="us-east-1")

response = client.invoke_model(
    modelId="us.anthropic.claude-opus-4-7",
    body=json.dumps({
        "anthropic_version": "bedrock-2023-05-31",
        "anthropic_beta": ["computer-use-2025-11-24"],
        "max_tokens": 2048,
        "messages": [
            {"role": "user", "content": "Take a screenshot of the desktop."}
        ],
        "tools": [
            {
                "type": "computer_20251124",
                "name": "computer",
                "display_width_px": 1024,
                "display_height_px": 768,
            }
        ],
    }),
)

result = json.loads(response["body"].read())
print(result["stop_reason"])  # "tool_use" when the model issues a computer action
print(result["content"])      # contains the tool_use block describing the action

The model requests the action through the computer tool:

{
    "type": "tool_use",
    "name": "computer",
    "input": {"action": "screenshot"}
}

InvokeModel with the computer toolset

The following example uses computer_toolset_20260801 instead. No beta value is required, and the toolset entry has no name.

import boto3
import json

client = boto3.client("bedrock-runtime", region_name="us-east-1")

response = client.invoke_model(
    modelId="us.anthropic.claude-opus-5-5",
    body=json.dumps({
        "anthropic_version": "bedrock-2023-05-31",
        "max_tokens": 2048,
        "messages": [
            {"role": "user", "content": "Take a screenshot of the desktop."}
        ],
        "tools": [{"type": "computer_toolset_20260801"}],
    }),
)

result = json.loads(response["body"].read())
print(result["content"])

The model calls the screenshot tool from the toolset directly. The toolset_name field identifies the toolset that the tool belongs to:

{
    "type": "tool_use",
    "name": "screenshot",
    "input": {},
    "toolset_name": "computer"
}

Converse

With the Converse API, pass the computer tool and the beta value in additionalModelRequestFields. The model's action is returned as a toolUse content block.

import boto3

client = boto3.client("bedrock-runtime", region_name="us-east-1")

response = client.converse(
    modelId="us.anthropic.claude-sonnet-4-6",
    messages=[
        {"role": "user", "content": [{"text": "Take a screenshot of the desktop."}]}
    ],
    inferenceConfig={"maxTokens": 2048},
    additionalModelRequestFields={
        "anthropic_beta": ["computer-use-2025-11-24"],
        "tools": [
            {
                "type": "computer_20251124",
                "name": "computer",
                "display_width_px": 1024,
                "display_height_px": 768,
            }
        ],
    },
)

print(response["stopReason"])  # "tool_use"
for block in response["output"]["message"]["content"]:
    if "toolUse" in block:
        print(block["toolUse"]["name"], block["toolUse"]["input"])

To use a client toolset with Converse, pass "tools": [{"type": "computer_toolset_20260801"}] in additionalModelRequestFields and omit anthropic_beta.

bedrock-mantle

On the bedrock-mantle endpoint (https://bedrock-mantle.region.api.aws/anthropic/v1/messages), send the beta value in the anthropic-beta header:

POST /anthropic/v1/messages
anthropic-version: 2023-06-01
anthropic-beta: computer-use-2025-11-24
content-type: application/json

{
    "model": "anthropic.claude-opus-5-5",
    "max_tokens": 2048,
    "messages": [
        {"role": "user", "content": "Take a screenshot of the desktop."}
    ],
    "tools": [
        {
            "type": "computer_20251124",
            "name": "computer",
            "display_width_px": 1024,
            "display_height_px": 768
        }
    ]
}

You can instead include "anthropic_beta": ["computer-use-2025-11-24"] in the request body. For authentication and the full bedrock-mantle Messages API request shape, see Inference using Anthropic Messages API.