Amazon Bedrock AgentCore Web Search integration

Amazon Bedrock AgentCore Web Search gives agents current information from the web with a source URL on every result. It is an AWS service-managed connector, so it authenticates with the AWS credentials you already have and takes no search API key of its own.

Overview

Integration details

Class	Package	Serializable	JS support	Version
`WebSearchToolkit`	`langchain-aws`	❌	❌

Tool features

Returns artifact	Native async	Return data	Pricing
❌	✅	Title, URL, Published date, Extract	Pay-per-use (AWS)

Available tools

Tool	Description
`web_search`	Search the web and return source-attributed results

Setup

The integration lives in the langchain-aws package, which wraps the bedrock-agentcore SDK.

pip install -U langchain-aws bedrock-agentcore

uv add langchain-aws bedrock-agentcore

Web search needs langchain-aws 1.7.9 or later and bedrock-agentcore 1.23.0 or later. Install bedrock-agentcore explicitly, as above. langchain_aws.tools imports the web search toolkit inside a try/except ImportError, so on an install without bedrock-agentcore, or with a version below 1.23.0, the export is skipped rather than raising, and create_web_search_toolkit is simply absent from langchain_aws.tools. If the import fails, check both installed versions before anything else.

Create a gateway with a web search target

Unlike the other AgentCore tools in this package, web search reaches the service through an AgentCore Gateway target, so a gateway has to exist before the toolkit can call it. You only do this once, and the same gateway serves every agent in the account.

from bedrock_agentcore.gateway.client import GatewayClient

client = GatewayClient(region_name="us-east-1")

gateway = client.create_gateway_and_wait(
    name="my-web-search-gateway",
    roleArn="arn:aws:iam::111122223333:role/MyGatewayExecutionRole",
    authorizerType="AWS_IAM",
    protocolType="MCP",
)

client.create_web_search_target(gateway_identifier=gateway["gatewayId"])

authorizerType="AWS_IAM" is what makes the caller’s own IAM permissions gate the gateway. NONE disables inbound authorization altogether, leaving the gateway callable by anyone who learns its URL and billed to your account. The target is named amazon-web-search by default, and Gateway prefixes every tool with its target name, so the tool on the gateway is amazon-web-search___WebSearch. The toolkit finds it for you; the name matters only if you go looking for it in the gateway’s tool list. Web search is available in us-east-1, eu-west-1, and ap-northeast-1. The service documentation carries the current list.

Credentials

Two different principals each need one permission, and mixing them up is the most common setup error:

Your own credentials need bedrock-agentcore:InvokeGateway on the gateway ARN.
The gateway’s execution role needs bedrock-agentcore:InvokeWebSearch on arn:aws:bedrock-agentcore:<region>:aws:tool/web-search.v1, and a trust policy that lets bedrock-agentcore.amazonaws.com assume it. It needs nothing else.

If that trust policy narrows aws:SourceArn to a region, it has to be the gateway’s region. A mismatch is accepted when the gateway is created and surfaces only on the first search, as Failed to obtain execution role credentials. It’s also helpful (but not needed) to set up LangSmith for best-in-class observability:

import os

os.environ["LANGSMITH_API_KEY"] = "your-api-key"
os.environ["LANGSMITH_TRACING"] = "true"

Instantiation

The factory is synchronous, unlike the browser and code interpreter factories in this package, because a search needs no session to be set up first.

from langchain_aws.tools import create_web_search_toolkit

toolkit, tools = create_web_search_toolkit(
    region="us-east-1",
    gateway_id="my-web-search-gateway-abc123",
)

Address the gateway by gateway_id, by gateway_arn, or by gateway_endpoint if you already know the MCP URL. A gateway_arn carries its own region. With a gateway_id, pass region to match the gateway’s region, or it will default to us-east-1. Passing target_name saves a tool discovery round trip on the first search. Use it if you named the target something other than the default:

toolkit, tools = create_web_search_toolkit(
    gateway_arn="arn:aws:bedrock-agentcore:us-east-1:111122223333:gateway/my-gw-abc123",
    target_name="amazon-web-search",
)

Invocation

Direct tool usage

search = tools[0]

results = search.invoke({"query": "latest boto3 release notes", "max_results": 5})
print(results)

Results come back as numbered text blocks, one per hit, each carrying a title, a URL, a publication date when the index reports one, and an extract:

1. boto3 1.43.56 release notes
   URL: https://github.com/boto/boto3/blob/develop/CHANGELOG.rst
   Published: 12:31PM, Friday, July 24 2026, PDT
   Adds support for ...

Use within an agent

from langchain.agents import create_agent
from langchain.chat_models import init_chat_model
from langchain_aws.tools import create_web_search_toolkit

toolkit, tools = create_web_search_toolkit(
    region="us-east-1",
    gateway_id="my-web-search-gateway-abc123",
)

llm = init_chat_model(
    "us.anthropic.claude-sonnet-4-20250514-v1:0",
    model_provider="bedrock_converse",
)

try:
    agent = create_agent(model=llm, tools=tools)
    result = agent.invoke(
        {"messages": [{"role": "user", "content": "What changed in the most recent boto3 release?"}]}
    )
    print(result["messages"][-1].content)
finally:
    toolkit.close()

The tool description asks the model to cite the URLs it used. Citing sources is a condition of use for this connector, so keep that instruction in place if you customize the prompt.

Filtering results

The model can narrow a search by domain and by publication date. Both are arguments on the tool, so the model can set them itself when the question calls for it:

# Only official documentation
search.invoke({
    "query": "S3 multipart upload limits",
    "include_domains": ["docs.aws.amazon.com"],
})

# Only recent pages
search.invoke({
    "query": "python release schedule",
    "published_after": "2026-01-01T00:00:00Z",
})

A root domain also matches its subdomains. include_domains can only narrow a search, never widen it: if the gateway target was created with its own include list, the two intersect, and disjoint lists return nothing. To enforce a domain policy that the agent cannot relax, set it on the target instead of leaving it to the model. Those lists are applied server-side and are not visible to the agent:

client.create_web_search_target(
    gateway_identifier=gateway["gatewayId"],
    include_domains=["docs.aws.amazon.com", "boto3.amazonaws.com"],
)

Error handling

A failed search raises ToolException, so a retry wrapper such as Runnable.with_retry() sees it as a failure rather than as a successful call that returned an error message. To let the model read the message and correct its own query instead, opt in per tool:

for tool in tools:
    tool.handle_tool_error = True

Releasing the client

The toolkit holds a client, which close() releases. It also works as a context manager:

toolkit, tools = create_web_search_toolkit(region="us-east-1", gateway_id="my-gw-abc123")

with toolkit:
    agent = create_agent(model=llm, tools=tools)
    ...

API reference

For detailed documentation of all features and configurations, see:

Connect these docs to your agent of choice via MCP for real-time answers.

Edit this page on GitHub or file an issue.

Popular Providers

Integrations by component

Amazon Bedrock AgentCore Web Search integration

Overview

Integration details

Tool features

Available tools

Setup

Create a gateway with a web search target

Credentials

Instantiation

Invocation

Direct tool usage

Use within an agent

Filtering results

Error handling

Releasing the client

API reference

​Overview

​Integration details

​Tool features

​Available tools

​Setup

​Create a gateway with a web search target

​Credentials

​Instantiation

​Invocation

​Direct tool usage

​Use within an agent

​Filtering results

​Error handling

​Releasing the client

​API reference

Overview

Integration details

Tool features

Available tools

Setup

Create a gateway with a web search target

Credentials

Instantiation

Invocation

Direct tool usage

Use within an agent

Filtering results

Error handling

Releasing the client

API reference