Does LangChain-AWS with Bedrock Claude 3.7 Sonnet support web search functionality?

[kytkhrz]

Hi there,

I recently came across an update on the Anthropic website stating that "Claude can now search the web" (source).
I wanted to confirm if this functionality is supported when using langchain-aws with the Bedrock Claude 3.7 Sonnet model.

If it is supported, could you provide guidance on how to enable or integrate the web search capability?
If not, are there plans to add this feature in the future?

Thank you for your help!

[Piyush Jain (3coins)]

kytkhrz
As far as I understand, this feature is in preview, and only available for the web version of Claude (claude.ai), not via the API, let me know if you can find more information about this support in Anthropic API.

[code-by-mahereddy]

As of now, web search functionality through Claude on AWS Bedrock is not supported in langchain-aws.

Here's why:

1. Anthropic's web search works with the Anthropic API directly, not through AWS Bedrock. It uses a server tool that Anthropic hosts.
2. AWS Bedrock is a managed service that proxies model calls. It doesn't expose all of Anthropic's native features (like web search), only what AWS has integrated.
3. The Bedrock Converse API does not have a web search tool capability in its current form.

Alternatives

If you need web search with LangChain, you can:

Option 1: Use LangChain's built-in web search tools

from langchain_community.tools import TavilySearchResults
from langchain_aws import ChatBedrockConverse

llm = ChatBedrockConverse(model="anthropic.claude-3-5-sonnet-20240620-v1:0")
search = TavilySearchResults()

# Bind the search tool to the LLM
llm_with_tools = llm.bind_tools([search])

Option 2: Use Anthropic API directly for web search

If you specifically need Anthropic's built-in web search, use ChatAnthropic from langchain-anthropic instead:

from langchain_anthropic import ChatAnthropic

llm = ChatAnthropic(model="claude-sonnet-4-20250514")
# Anthropic's web search is available via their API

Option 3: Implement a custom tool

from langchain_core.tools import tool
import requests

@tool
def web_search(query: str) -> str:
    "Search the web for information."
    # Use your preferred search API
    response = requests.get(f"https://api.search.service/search?q={query}")
    return response.text

llm_with_tools = llm.bind_tools([web_search])

For the latest updates on Bedrock feature support, check the AWS Bedrock documentation.

[Lucas Kim (kimnamu)]

Thanks kytkhrz for linking the Anthropic announcement in the question, and Piyush Jain (@3coins) and code-by-mahereddy for the earlier answers. The core point still holds and I re-checked it below: Anthropic's server-side web_search tool does not reach Claude on Bedrock, directly or through ChatBedrockConverse.

What has changed since June is the second half. langchain-aws 1.7.9 (2026-09-22) ships its own web search tool, create_web_search_toolkit, backed by Web Search on Amazon Bedrock AgentCore (#1275), plus a sample agent notebook, samples/agents/agentcore_web_search_agent.ipynb (#1280). It returns a regular LangChain tool, so it binds to ChatBedrockConverse like any other, and calls are signed with your AWS credentials; there is no search API key.

Path	Checked with	Result
Anthropic web_search_20250305 via InvokeModel	Bedrock us-east-1, Haiku 4.5	ValidationException: tool type 'web_search_20250305' is not supported for this model
Same tool via Converse additionalModelRequestFields	Bedrock us-east-1, Haiku 4.5	ValidationException, same message
ChatBedrockConverse.bind_tools([{"type": "web_search_20250305", ...}])	langchain-aws 1.7.9	KeyError: 'parameters' in _format_tools (L3526); only client-side tool specs are accepted
create_web_search_toolkit(region=..., gateway_id=..., target_name=...)	langchain-aws 1.7.9, bedrock-agentcore 1.23.1, offline	a web_search tool with query, max_results, include_domains, exclude_domains, published_after, published_before

# pip install "langchain-aws[tools]"      # brings bedrock-agentcore>=1.23.0
from langchain.agents import create_agent
from langchain_aws import ChatBedrockConverse
from langchain_aws.tools import create_web_search_toolkit

toolkit, tools = create_web_search_toolkit(
    region="us-east-1",                 # us-east-1, eu-west-1 or ap-northeast-1, per the notebook
    gateway_id="<AgentCore gateway with a web-search target>",
    target_name="amazon-web-search",
)
model = ChatBedrockConverse(model="us.anthropic.claude-sonnet-4-5-20250929-v1:0")
agent = create_agent(model, tools=tools)

Two caveats. The toolkit needs an AgentCore Gateway with the web search connector target, and the connector is enabled per account; the notebook creates the gateway and lists the two IAM permissions involved. I had no gateway to point at, so the last row is the offline construction plus the source, not a live search. On the model: anthropic.claude-3-7-sonnet-* is no longer in list-foundation-models for Anthropic in us-east-1 on my account, so the notebook uses a current cross-region profile and says any Bedrock model you have access to works.

Run output and source pointers

Bedrock and ChatBedrockConverse with the Anthropic server tool:

langchain-aws 1.7.9 / boto3 1.43.100 / model us.anthropic.claude-haiku-4-5-20251001-v1:0
A invoke_model -> ValidationException : tool type 'web_search_20250305' is not supported for this model
B converse -> ValidationException : The model returned the following errors: tool type 'web_search_20250305' is not supported for this model
C bind_tools -> KeyError : 'parameters' | raised at langchain_aws/chat_models/bedrock_converse.py L 3526 _format_tools

The new toolkit, constructed with sockets blocked (no gateway available to me):

langchain-aws 1.7.9 / bedrock-agentcore 1.23.1
tool: web_search
args: ['query', 'max_results', 'include_domains', 'exclude_domains', 'published_after', 'published_before']
description head: Search the web for current information and return source-attributed results.

Source, main 2e49945: libs/aws/langchain_aws/tools/web_search_toolkit.py (create_web_search_toolkit at L366, added by #1275, included in tag langchain-aws==1.7.9); the tools extra is libs/aws/pyproject.toml L26 (bedrock-agentcore>=1.23.0); _format_tools converts every tool through convert_to_openai_tool and reads spec["parameters"] at bedrock_converse.py L3526, which is where a server-tool dict without an input schema fails. The regions, the per-account enablement and the IAM permissions are from the first two markdown cells of samples/agents/agentcore_web_search_agent.ipynb.