On Anthropic's first-party API, you can hand Claude a server-side web search tool and let Anthropic's infrastructure run the searches. On Amazon Bedrock, none of the server-side web tools exist: web search, web fetch, and code execution are all unsupported there. What Bedrock does fully support is standard tool use (sometimes called function calling) — the mechanism where you describe a tool in your request, Claude decides when to call it, and your code executes the call and returns the result. That is the entire basis of the workaround: bring your own search backend and expose it to Claude as a tool.
The architecture in one paragraph
Your application defines a web_search tool with a simple input schema (a query string). When a user asks something that needs fresh information, Claude responds with a tool-use request instead of an answer. Your code runs the query against a search backend you operate or subscribe to — a commercial search API, an internal index, or both — then sends the results back as a tool result. Claude reads them and writes the final answer, grounded in what the search returned. Nothing about this requires Bedrock-specific features beyond ordinary tool use.
A minimal implementation
from anthropic import AnthropicBedrockMantle
client = AnthropicBedrockMantle(aws_region="us-east-1")
tools = [{
"name": "web_search",
"description": "Search the web for current information.",
"input_schema": {"type": "object",
"properties": {"query": {"type": "string"}},
"required": ["query"]},
}]
messages = [{"role": "user", "content": "What changed in our industry this week?"}]
response = client.messages.create(model="anthropic.claude-sonnet-5",
max_tokens=2048, tools=tools,
messages=messages)
# If response.stop_reason == "tool_use": run the search yourself,
# append a tool_result message with the findings, and call again.The loop after the comment is the part you own: extract the tool-use block, call your search backend, and append a tool_result content block with the findings before invoking the model again. Most production versions run this loop until Claude stops asking for tools.
Feeding results back well
Result formatting matters more than backend choice. Return a handful of results with title, URL, and a text snippet — enough for Claude to judge relevance and cite sources, not entire raw pages. Claude on Bedrock supports search-results content blocks and citations, which let you return results in a structured form and get answers that attribute claims to specific sources; that combination is worth using if verifiable answers are a requirement. If you need full-page content, fetch and extract the text in your own code first, and trim it — every character you return is input tokens you pay for.
What you take on by self-managing
| Concern | Who handles it in this pattern |
|---|---|
| Search quality and coverage | Your chosen backend and your query handling |
| Search cost and rate limits | Your contract with the search provider |
| Fetching and sanitizing page content | Your code — treat fetched text as untrusted input |
| Compliance and egress review | Your security team; searches leave your boundary to whatever backend you pick |
The last two rows deserve emphasis. Content retrieved from the open web can contain adversarial instructions aimed at your model (prompt injection), so treat search results as data, never as trusted instructions — constrain what actions Claude can take based on them, and prefer read-only tools in this loop. And because your application, not the model platform, calls the search provider, that egress is a new data flow for your security review: what queries leave, to whom, and what gets logged.
One more option check before you build: if your organization also runs Claude Platform on AWS (Anthropic-operated, inside AWS), the server-side web search and web fetch tools are available there. For teams committed to Bedrock proper, the tool-use pattern above is the supported path.
Where to go next
See the Bedrock feature availability matrix for the full list of gaps, prompt injection 101 before you wire untrusted web content into prompts, and web tools on Claude Platform on AWS if you want the managed alternative.