llama-index-tools-aws-bedrock-agentcore 0.3.0

llama-index Amazon Bedrock AgentCore Runtime and Tools integration

Amazon Bedrock AgentCore Runtime and Tools

This module provides a runtime adapter and tools for deploying and extending LlamaIndex agents with Amazon Bedrock AgentCore -- including managed compute via AgentCore Runtime, sandboxed browser automation, and code execution.

Prerequisites

• AWS credentials configured via environment variables, AWS CLI profile, or IAM role
• IAM permissions for bedrock-agentcore:* actions (see the AgentCore documentation for details)
• Python 3.9+

Installation

(Optional) To run the examples below, first install:

pip install llama-index llama-index-llms-bedrock-converse

Install the main tools package:

pip install llama-index-tools-aws-bedrock-agentcore

Runtime

The AgentCoreRuntime adapter deploys any LlamaIndex agent to Amazon Bedrock AgentCore Runtime -- a managed compute platform for AI agents. It wraps BedrockAgentCoreApp from the bedrock-agentcore SDK, providing the required POST /invocations and GET /ping endpoints.

Quick Start

from llama_index.llms.bedrock_converse import BedrockConverse
from llama_index.core.agent.workflow import FunctionAgent
from llama_index.tools.aws_bedrock_agentcore import AgentCoreRuntime

llm = BedrockConverse(
    model="us.anthropic.claude-sonnet-4-6-v1",
    region_name="us-west-2",
)
agent = FunctionAgent(llm=llm, tools=[])

# One-liner -- starts uvicorn on port 8080
AgentCoreRuntime.serve(agent)

With Options

runtime = AgentCoreRuntime(
    agent=agent,
    stream=True,  # SSE streaming (default)
    port=8080,  # Required port for AgentCore deployment
    debug=False,
)
runtime.run()

With AgentCore Memory

from llama_index.memory.bedrock_agentcore import (
    AgentCoreMemory,
    AgentCoreMemoryContext,
)

memory = AgentCoreMemory(
    context=AgentCoreMemoryContext(
        memory_id="your-memory-id",
        actor_id="user-123",
    ),
    region_name="us-west-2",
)

# Session ID from the X-Amzn-Bedrock-AgentCore-Runtime-Session-Id header
# is automatically wired to memory
AgentCoreRuntime.serve(agent, memory=memory)

Sending Requests

# Non-streaming
curl -X POST http://localhost:8080/invocations \
  -H "Content-Type: application/json" \
  -d '{"prompt": "Hello, what can you do?"}'

# Streaming (SSE)
curl -N -X POST http://localhost:8080/invocations \
  -H "Content-Type: application/json" \
  -d '{"prompt": "Hello, what can you do?"}'

The adapter accepts prompt, message, or input as the payload key.

Streaming Event Types

When stream=True (default), the SSE stream emits these event types:

Event	Fields	Description
agent_stream	delta, response, thinking_delta?	Token-by-token LLM output
tool_call	tool_name, tool_kwargs	Before tool execution
tool_result	tool_name, tool_output	After tool execution
done	response	Final agent response
error	message	Error during streaming

Testing with ASGI

runtime = AgentCoreRuntime(agent=agent)
app = runtime.app  # BedrockAgentCoreApp (Starlette-based)
# Use with httpx.AsyncClient for testing

Toolspecs

Browser

The AgentCore Browser toolspec provides a set of tools for interacting with web browsers in a secure sandbox environment. It enables your LlamaIndex agents to navigate websites, extract content, click elements, and more.

Included tools:

• navigate_browser: Navigate to a URL
• click_element: Click on an element using CSS selectors
• extract_text: Extract all text from the current webpage
• extract_hyperlinks: Extract all hyperlinks from the current webpage
• get_elements: Get elements matching a CSS selector
• navigate_back: Navigate to the previous page
• current_webpage: Get information about the current webpage
• generate_live_view_url: Generate a presigned URL for human oversight of a browser session
• take_control: Take manual control of a browser session (disables automation)
• release_control: Release manual control (re-enables automation)

Lifecycle methods available for programmatic use (not exposed as agent tools):

• list_browsers, create_browser, delete_browser, get_browser

You can optionally pass a custom identifier for VPC-enabled browser resources:

tool_spec = AgentCoreBrowserToolSpec(
    region="us-west-2",
    identifier="my-custom-browser-id",
)

Example usage:

import asyncio
from llama_index.llms.bedrock_converse import BedrockConverse
from llama_index.tools.aws_bedrock_agentcore import AgentCoreBrowserToolSpec
from llama_index.core.agent.workflow import FunctionAgent

import nest_asyncio

nest_asyncio.apply()  # In case of existing loop (ex. in JupyterLab)


async def main():
    tool_spec = AgentCoreBrowserToolSpec(region="us-west-2")
    tools = tool_spec.to_tool_list()

    llm = BedrockConverse(
        model="us.anthropic.claude-sonnet-4-6-v1",
        region_name="us-west-2",
    )

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

    task = "Go to https://news.ycombinator.com/ and tell me the titles of the top 5 posts."

    response = await agent.run(task)
    print(str(response))

    await tool_spec.cleanup()


if __name__ == "__main__":
    asyncio.run(main())

Code Interpreter

The AgentCore Code Interpreter toolspec provides a set of tools for interacting with a secure code interpreter sandbox environment. It enables your LlamaIndex agents to execute code, run shell commands, manage files, and perform computational tasks.

Included tools:

• execute_code: Run code in various languages (primarily Python)
• execute_command: Run shell commands
• read_files: Read content of files in the environment
• list_files: List files in directories
• delete_files: Remove files from the environment
• write_files: Create or update files
• start_command: Start long-running commands asynchronously
• get_task: Check status of async tasks
• stop_task: Stop running tasks
• upload_file: Upload a file with an optional semantic description
• upload_files: Upload multiple files at once
• install_packages: Install Python packages via pip
• download_file: Download a file from the sandbox
• download_files: Download multiple files from the sandbox
• clear_context: Clear all variable state in the Python execution context

Lifecycle methods available for programmatic use (not exposed as agent tools):

• list_code_interpreters, create_code_interpreter, delete_code_interpreter, get_code_interpreter

You can optionally pass a custom identifier for VPC-enabled code interpreter resources:

tool_spec = AgentCoreCodeInterpreterToolSpec(
    region="us-west-2",
    identifier="my-custom-interpreter-id",
)

Example usage:

import asyncio
from llama_index.llms.bedrock_converse import BedrockConverse
from llama_index.tools.aws_bedrock_agentcore import (
    AgentCoreCodeInterpreterToolSpec,
)
from llama_index.core.agent.workflow import FunctionAgent

import nest_asyncio

nest_asyncio.apply()  # In case of existing loop (ex. in JupyterLab)


async def main():
    tool_spec = AgentCoreCodeInterpreterToolSpec(region="us-west-2")
    tools = tool_spec.to_tool_list()

    llm = BedrockConverse(
        model="us.anthropic.claude-sonnet-4-6-v1",
        region_name="us-west-2",
    )

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

    code_task = "Write a Python function that calculates the factorial of a number and test it."

    code_response = await agent.run(code_task)
    print(str(code_response))

    command_task = "Use terminal CLI commands to: 1) Show the environment's Python version. 2) Show me the list of Python package currently installed in the environment."

    command_response = await agent.run(command_task)
    print(str(command_response))

    await tool_spec.cleanup()


if __name__ == "__main__":
    asyncio.run(main())

Example Notebooks

• Browser Tool Notebook
• Code Interpreter Tool Notebook