Skip to main content

Amazon Bedrock

Web Search

Connect your AI agents and applications to the web. Cited results you can trace, with zero data egress by default.

Benefits

Why teams reach for Web Search instead of a third-party provider.

    By default, Web Search offers zero data egress, so your data never leaves your AWS environment. Queries are served from an Amazon-operated web index with no third-party search provider in the request path, and processing is strictly in-Region with no cross-Region routing. Web Search inherits the Amazon Bedrock compliance posture, including SOC 1/2/3, ISO 27001, HIPAA eligibility, FedRAMP High, and PCI DSS, which removes the separate security and legal review a third-party provider would require.

    Web Search uses your existing AWS credentials, so there are no new API keys to provision and no separate SDK to adopt. It becomes a single parameter on the OpenAI-compatible call you already make, with no function schema to define and no client-side tool-use loop, which removes the vendor onboarding, contracts, and key management a third-party provider would need before the first query ever runs. On AgentCore it is a managed MCP connector on Gateway, reachable from any MCP-compatible client, including Strands Agents, LangChain, LangGraph, and CrewAI.

    Web Search is backed by a web index Amazon operates, spanning billions of documents and refreshed continually. It combines that index with a built-in Amazon Knowledge Graph that anchors the entities in a domain along with the connections between them. Where a question has a definite answer, the knowledge graph supplies it with strong confidence rather than leaving the model to reconstruct it from page fragments. The same infrastructure powers Alexa+, Kiro, and Amazon Quick.

    Results come back with the snippet plus the source details an agent needs to attribute it, such as the page URL, title, and publication date. Web Search returns the material; the model or agent decides what to cite and how to display it.

Customers

How teams are putting Web Search into production.
Missing alt text value
At Halo Ai, data is the foundation of our risk intelligence platform. Amazon AgentCore Web Search gives us access to timely public information that helps enrich the data feeding our intelligence, reasoning, and continuous monitoring capabilities. This helps us deliver on our customer promise of Verified Trust at Scale.

Nagy Ghobrial

CTO, Halo AI
Prox logo
Prox's autonomous research agents use Web Search on Amazon Bedrock AgentCore to understand what buyers are saying and how competitors position their alternatives. In July, Prox used Web Search across more than 5,800 research sessions and nearly 75,000 successful search calls, with 99.8% completing in under one second.

Dima Yanovsky

Founder and CEO, Prox Inc
Exaforce logo
Exaforce is built entirely on AWS, so enriching a detection with reputation data on the IPs and domains involved meant stepping outside it, where every lookup cost us money and time. Web Search on Amazon Bedrock put that step back inside our own stack and eliminated the vendor contract, the external integration and the added latency.

Jakub Pavlik

Co-Founder and Head of Engineering, Exaforce

FAQs

Web Search

Open all

    On Amazon Bedrock, Web Search runs on OpenAI GPT-5.4 and later models, served through the bedrock-mantle endpoint using the Responses API. On AgentCore there is no model restriction: Web Search is a managed MCP connector on AgentCore Gateway, so your agent retrieves the results and grounds any model with them, from any MCP-compatible client including Strands Agents, LangChain, LangGraph, and CrewAI.

    By default, no, and the detail differs by surface. On AgentCore, queries are served entirely within AWS infrastructure: they are never sent to a third-party search engine or routed outside AWS, and the data path stays inside AWS end to end. On Amazon Bedrock, search results always come from the web index. External web access, also called live fetch, is what lets the tool read a page directly from the public web, and it is off by default. Turning it on takes a deliberate change on your side, because it requires an additional IAM permission that the standard Amazon Bedrock policies do not grant. Until you make that change, retrieval stays within the index and cache and your request data does not leave the AWS boundary.

    Web Search does not persist your prompt: neither your input nor the search queries the model generates are written into durable storage. Search results, including titles, source URLs, and snippets, are not stored against your account, your request identity, or your prompt text, and your prompts and tool calls are not sent to a third-party search provider. Depending on the model you use, your request may still be subject to automated Amazon Bedrock abuse detection, and retention of prompts and model outputs is governed by your account or project data retention mode.

    Web Search is integrated with AWS CloudTrail, which records each search as a data event with the calling identity, timestamp, action, source identity, and the account and Region context of the request. Data events are not logged by default: create a trail or event data store and enable data event logging for Web Search. By design, CloudTrail does not record query text or the URLs search returned, so you get a full record of who used the tool and when without exposing what end users searched for.

Did you find what you were looking for today?

Let us know so we can improve the quality of the content on our pages