My Motivation
As AI adoption grows inside organizations, both employee AI sessions and autonomous agents need access to the same business context and tools.
Amazon Bedrock AgentCore allows us to centrally host modular MCP servers and make them available across both execution models. This gives us a single place to improve and maintain each set of capabilities while hundreds of sessions and agents benefit from every change, reducing costs and helping employees work more efficiently.
A real-world use case
Before introducing these MCP servers, every employee had to provide the required context in each new AI session.
For support workflows, users had to explain the ticket, identify the relevant codebase, and repeatedly describe the conventions for comments, status updates, and pull requests. Without shared context, different sessions could search the code differently or produce changes that did not follow the team’s established patterns.
The same problem existed when accessing the lakehouse. Employees and agents had to determine which tables to query, how to query them, and how each business metric was calculated in SQL. This context was reconstructed independently in every session.
We addressed this by creating modular MCP servers for each business context. One provides the context and tools required for support and development workflows, while another exposes the knowledge and capabilities needed to work consistently with our lakehouse.
The solution: Hosting our shared MCP layer on Amazon Bedrock AgentCore
To make these modular capabilities reusable across the organization, we needed a central place to run them. Amazon Bedrock AgentCore Runtime allows us to host MCP servers in a serverless environment and make the same endpoints available to both employee AI sessions and autonomous agents.
AgentCore also supports centralized authentication and authorization, CloudWatch logs and operational metrics, and the ability to add detailed tracing through OpenTelemetry.
This turns our MCP servers from locally configured tools into a shared organizational layer that can be improved, monitored, and secured from one place.
The cost impact
This shared MCP layer also changes how we think about cost. Employee-driven tasks can use existing organizational AI subscriptions, while autonomous workflows continue to run on AgentCore and consume models through Amazon Bedrock.
For example, at current pricing, aggregate usage of 100,000 input tokens and 10,000 output tokens with Claude Sonnet 4.6 costs approximately $0.45 in Bedrock inference, or around $45 for 100 similar tasks, excluding AgentCore Runtime charges and, if used, AgentCore Gateway charges.
If employees already have AI subscriptions with included usage, interactive tasks can use that existing capacity—subject to plan limits—instead of generating additional pay-per-token inference costs.
The goal is not to replace autonomous agents. It is to reserve pay-per-token execution for tasks that require automation, while employee-driven work can use existing AI sessions. MCP makes this possible without duplicating business context or tools.
Loading article