Amazon Bedrock AgentCore Web Search: Pricing, Setup, and Risks

Last updated: 2026-07-23

Last verified: 2026-07-23

TL;DR: Amazon Bedrock AgentCore Web Search is a managed, MCP-compatible search connector for agents using AgentCore Gateway. AWS says queries stay within AWS infrastructure and results include current snippets, source URLs, titles, and publication dates. The service launched at $7 per 1,000 searches in US East (N. Virginia), with Gateway and related AWS costs considered separately.

What launched?

AWS announced general availability in June 2026. Developers can add Web Search as a target on an AgentCore Gateway, let an agent discover it through the standard MCP tool flow, and receive structured web observations without provisioning a separate search API or storing third-party search credentials.

AWS says the connector uses an Amazon-operated index spanning tens of billions of documents, refreshed continually, with semantic snippet extraction and a knowledge graph for some entity queries. Those are service claims that still need workload-specific evaluation for relevance, coverage, and freshness.

How the connector works

The gateway target uses the connector ID web-search. The application authenticates to AgentCore Gateway, which invokes the AWS-owned Web Search tool. The tool returns a serialized set of observations with text, title, URL, and publication date fields that the agent can use when composing an answer.

This is public-web grounding, not enterprise document retrieval. AWS positions Bedrock Knowledge Bases for private organizational data and Web Search for information that changes outside the organization. Teams following related AI News should keep those retrieval paths and their access controls separate.

Where it fits

Research agents that need current product, policy, market, or release information.
Customer-support assistants that must cite changing public documentation.
AWS-governed agent systems that do not want third-party search credentials.
MCP-compatible applications already routing tools through AgentCore Gateway.
Workflows that combine public-web search with a separate private knowledge base.

Pricing and availability

AWS lists Web Search at $7 per 1,000 queries with no upfront commitment. The launch material says the connector is available in us-east-1. Gateway invocations, models, logs, data transfer, and other AgentCore resources may add cost, so the search price is only one line in a production estimate.

There is no standalone free-plan promise for the connector. New AWS customers may have general account credits, but credits are not the same as a permanent Web Search tier. Check the current AgentCore pricing page before committing traffic.

Setup outline

Create or select an AgentCore Gateway in a supported Region.
Add a target with connectorId set to web-search.
Grant the gateway role the documented InvokeGateway and InvokeWebSearch permissions.
Discover the tool through MCP and run a small set of known-answer queries.
Log returned sources, latency, empty results, and downstream model citations.

What to measure

Evaluate source quality, long-tail coverage, freshness, citation accuracy, latency, and the number of searches an agent issues per task. Agent loops can call tools repeatedly, so a nominal per-query price can grow if prompts or orchestration encourage redundant searches.

Test queries that should fail as well as those that should succeed. A good connector needs predictable behavior for obscure topics, conflicting sources, missing dates, unsafe pages, and queries that include sensitive customer information.

Risks and controls

Public webpages can contain errors, adversarial instructions, copied content, and prompt injection. Keeping the query path inside AWS does not make every retrieved page trustworthy. The application should separate retrieved text from system instructions, constrain tool permissions, and retain the source list for review.

Regional concentration may also matter for architecture and compliance. Confirm where the gateway, model, logs, and search connector operate, then document what data is sent in each query. Avoid putting private records into a public-web search request.

Alternatives

Alternatives include other managed grounding services, specialist search APIs, self-operated crawlers, and retrieval pipelines built over approved sources. AgentCore Web Search is most attractive when the application already uses AWS identity, AgentCore Gateway, and MCP tooling. A standalone search API may be simpler for a product outside that stack.

Kingy AI verdict

Test the connector if an AWS-based agent needs recent public information and the team wants search governed through AgentCore. Start with a capped evaluation workload. Measure citation quality and query count before assuming the managed path is cheaper or more reliable than the current search integration.

FAQ

How much does AgentCore Web Search cost?

AWS announced a price of $7 per 1,000 search queries. Other AgentCore and AWS usage may be billed separately.

Does the query leave AWS?

AWS says Web Search queries are served within its infrastructure and are not sent to a third-party search engine.

Is Web Search the same as a Bedrock Knowledge Base?

No. Web Search covers the public web. Knowledge Bases are designed for content an organization ingests and controls.