AWS Opens Direct Web Search API at $7 per 1,000 Queries

Amazon Web Services announced a standalone Web Search API for Amazon Bedrock AgentCore on November 11, 2026, priced at $7 per 1,000 queries. One signed call now returns ranked web results. No gateway to build, no separate API key, no third-party search engine in the loop.

That last point matters. AWS says queries run inside AWS against an Amazon-operated web index — the same search infrastructure behind Alexa+, Kiro, Quick and other Amazon products. For enterprises building agents and data pipelines, that means retrieval traffic stays on Amazon's network instead of flowing to an outside search vendor. In a market where grounding AI answers in fresh web data is table stakes, AWS just made its index programmatically reachable with minimal ceremony.

What changed on November 11?

Before this launch, web search on AgentCore ran exclusively through AgentCore Gateway. That path required three pieces: a gateway, a web search target, and a Model Context Protocol (MCP) client. Developers assembling custom stacks often found that overhead unnecessary.

The new InvokeWebSearch operation replaces it with a plain API call. It works through standard AWS SDKs or the AWS CLI, authenticating with ordinary AWS Identity and Access Management (IAM) credentials and Signature Version 4. There's no separate API key or secret to manage.

Results come back with text extracts and, when available, page titles, URLs and publication times.

Who is the direct API for?

AWS pitches the standalone API at two audiences. The first is developers running custom agent loops and frameworks — LangChain and LlamaIndex get named explicitly. The second is non-agent workloads entirely:

• Dashboards
• Fact-checking pipelines
• Content enrichment
• Monitoring systems

The API also hands operational control back to the caller. Developers manage retries, caching and error handling themselves rather than delegating those functions to a gateway. For teams with opinionated infrastructure, that trade — more code for more control — is often worth making.

What are the technical limits?

The constraints are concrete and documented. Queries can run up to 200 characters. Callers request between 1 and 25 results, and receive up to 25 when maxResults is omitted. Domain include and exclude lists each hold as many as 100 entries, and specifying a root domain also matches its subdomains.

Publication filters accept UTC timestamps in RFC 3339 format, with both publishedAfter and publishedBefore bounds inclusive.

On the tooling side, Python callers need Boto3 1.43.127 or later. Support also ships in langchain-aws 1.9.3 and the AgentCore Python SDK 1.26.0. The AgentCore console lets users test queries with result-count, date and domain settings, inspect result metadata and call latency, then generate an integration snippet or create a gateway target.

How do identity and network controls work?

Permission is straightforward for anyone already inside the AWS ecosystem. Callers need bedrock-agentcore:InvokeWebSearch permission on the AWS-managed web search tool — the same action and resource a gateway's service role uses. Service control policies can apply guardrails across accounts, letting organizations govern direct search alongside their other AWS permissions.

Private networking gets first-class treatment. For workloads in private subnets, AWS PrivateLink routes calls through an interface endpoint in a virtual private cloud. Traffic stays on the AWS network with no internet gateway, NAT gateway or proxy required. Endpoint policies, security groups, subnet placement and an aws:SourceVpce IAM condition can restrict which principals and network paths may invoke search.

What about audit trails and compliance?

CloudTrail records InvokeWebSearch requests as data events, but — a wrinkle worth noting — these are not logged by default. A trail or event data store needs an explicit event selector for the AWS::BedrockAgentCore::WebSearch resource type. The resulting events identify the web search tool and maxResults, but not the search text or filters.

One contractual requirement stands out: the acceptable-use terms oblige customers to retain and display the source citations and links supplied with results. Attribution isn't optional.

Where is it available, and what does it cost?

The API meters to the same usage type as the gateway connector but incurs no gateway charges on the direct path. Billing stays at $7 per 1,000 queries.

The default quota is 10 requests per second per account in each Region, and customers can request an increase. At launch, availability covers six Regions:

• US East (N. Virginia)
• US East (Ohio)
• US West (Oregon)
• Europe (Ireland)
• Asia Pacific (Tokyo)
• Asia Pacific (Sydney)

When should you still use the gateway?

AWS isn't retiring the gateway, and it still recommends AgentCore Gateway for most agentic workloads. The recommendation is strongest for deployments that combine MCP-compatible tools behind one endpoint, use optional MCP sessions, or apply a single administrator-controlled source policy across teams.

Both routes search the same index, but their security models differ. The gateway can authorize callers with JSON Web Tokens; the standalone API uses IAM only. Gateway administrators can also enforce target-level domain lists hidden from calling agents, while direct API requests carry their own filters in the open.

The dividing line AWS has drawn is operational control. Direct callers own request handling and filtering themselves. Shared agent environments keep the gateway's centralized tool, authentication and source-policy controls.

The strategic significance is that Amazon's web index is now accessible without adopting MCP or the full gateway stack. That lowers the barrier for data teams and framework builders who want grounded search results inside AWS. With its metering identical to the gateway path and PrivateLink support built in, the direct API positions AWS to capture the retrieval layer of agentic workloads it previously forced through a heavier architectural door.