Amazon Bedrock AgentCore Browser Tools

This notebook demonstrates how to use the Amazon Bedrock AgentCore Browser tools with LlamaIndex. The toolkit provides a managed Chrome browser for web automation, content extraction, and human-in-the-loop oversight via live view URLs.

Prerequisites

• An AWS account with Amazon Bedrock AgentCore access
• Configured AWS credentials (via ~/.aws/credentials, environment variables, or IAM role)
• Required IAM permissions: bedrock-agentcore:InvokeBrowser
• Playwright is installed automatically as a dependency of the toolkit

Install Dependencies

%pip install -q llama-index llama-index-tools-aws-bedrock-agentcore llama-index-llms-bedrock-converse

Setup

Create an AgentCoreBrowserToolSpec instance. A browser session is started lazily on first navigation.

from llama_index.tools.aws_bedrock_agentcore import AgentCoreBrowserToolSpec

tool_spec = AgentCoreBrowserToolSpec(region="us-west-2")
tools = tool_spec.to_tool_list()

print(f"Available tools: {[t.metadata.name for t in tools]}")

Direct Tool Usage

You can call the tool spec methods directly without an agent. Since Jupyter runs inside an asyncio event loop, use the async methods (prefixed with a).

Navigate and Extract Content

# Navigate to a page (use async methods inside Jupyter's event loop)
result = await tool_spec.anavigate_browser("https://example.com")
print(result)

# Get current page info
result = await tool_spec.acurrent_webpage()
print(result)

# Extract text from a specific element
result = await tool_spec.aextract_text(selector="h1")
print(f"Page heading: {result}")

Links and Elements

# Extract all hyperlinks
result = await tool_spec.aextract_hyperlinks()
print("Hyperlinks:")
print(result)

# Get elements matching a CSS selector
result = await tool_spec.aget_elements("p")
print("\nParagraph elements:")
print(result)

Live View URL

Generate a presigned URL that lets a human observe the browser session in real time. This is useful for oversight and debugging.

live_url = await tool_spec.agenerate_live_view_url()
print(f"Live view URL (expires in 5 minutes):\n{live_url}")

Human-in-the-Loop Control

Use take_control() to pause automation and let a human interact with the browser via the live view URL. Call release_control() to hand control back to the agent.

# Pause automation for human interaction
result = await tool_spec.atake_control()
print(result)

# ... human interacts via the live view URL ...

# Resume automation
result = await tool_spec.arelease_control()
print(result)

Agent Integration

Combine the browser tools with a FunctionAgent and BedrockConverse LLM to let the agent autonomously browse the web.

from llama_index.llms.bedrock_converse import BedrockConverse
from llama_index.core.agent.workflow import FunctionAgent

llm = BedrockConverse(model="us.anthropic.claude-sonnet-4-20250514-v1:0")

agent = FunctionAgent(
    tools=tools,
    llm=llm,
)

Web Scraping Task

Ask the agent to visit Hacker News and summarize the top posts.

response = await agent.run(
    "Go to https://news.ycombinator.com/ and tell me the titles of the top 5 posts."
)
print(str(response))

Content Extraction Task

Ask the agent to visit a page and extract specific information.

response = await agent.run(
    "Go to https://example.com and extract the page heading and all hyperlinks on the page."
)
print(str(response))

Cleanup

Always clean up browser sessions when done to release resources.

await tool_spec.cleanup()
print("Browser sessions cleaned up.")

References

• Amazon Bedrock AgentCore Documentation
• llama-index-tools-aws-bedrock-agentcore on PyPI