HEMA replaced a patchwork of wikis, service catalogs, and IT portals with a single knowledge layer called HAL, built on Model Context Protocol (MCP) and Amazon Bedrock AgentCore. The shift eliminates manual portal‑hopping, delivers answers inside developers’ existing tools, and centralises governance under Microsoft Entra ID.
Why the Change Matters to Practitioners
Engineers, SREs, and security staff all spend time locating procedural information – API access requests, group provisioning steps, or policy clarifications. By exposing that knowledge through a unified layer, teams can retrieve answers in seconds from an IDE chat window or a familiar conversational agent. The result is faster onboarding, reduced context‑switching, and a consistent source of truth that scales as the organisation grows.
Knowledge Layer with MCP: Design Overview
The solution hinges on two separable responsibilities:
- HAL (the knowledge layer) aggregates structured data from HEMA’s service catalog, product‑information‑management system, and data‑mesh tables into a single, governed repository.
- MCP provides a uniform API that any client – the HAL web chat, Kiro, Claude, or future agents – can call to retrieve that knowledge without bespoke integrations.
Amazon Bedrock AgentCore hosts the HAL agent, handling request routing, model selection, and scaling without requiring HEMA to operate dedicated MCP server infrastructure. Each knowledge source is registered once as an MCP‑compatible tool, and the same endpoint serves all downstream clients.
Operational and Security Considerations
Security is anchored in Microsoft Entra ID; client applications never hold AWS credentials, limiting the attack surface. Authentication and authorization are therefore delegated to the corporate identity provider, while Bedrock AgentCore manages the model execution environment.
- Operational teams must monitor Bedrock usage metrics to anticipate scaling needs as more tools adopt the HAL chat.
- Because the knowledge layer is read‑only today, change management processes remain unchanged – updates to the service catalog flow directly into HAL.
- Future extensions toward an “action layer” will introduce write‑back capabilities, requiring additional governance and audit controls.
Related CloudNinjas coverage: AWS.
What This Means For Practitioners
Adopting a knowledge layer with MCP and Bedrock AgentCore offers a pragmatic path to eliminate fragmented documentation and embed answers in daily workflows. Practitioners should evaluate the effort to expose existing data sources as MCP tools, verify that their identity provider can supply the required Entra ID integration, and plan for operational monitoring of Bedrock workloads. When the read‑only model proves stable, the same pattern can be extended to trigger automated actions, but that step will demand tighter security reviews and change‑control processes.
