HEMA faced a classic knowledge-fragmentation problem: engineers had to navigate multiple portals, wikis, and service catalogs to answer routine questions like "How do I request API access?" or "Who owns this service?" (AWS Machine Learning Blog). The retailer built HAL, an internal AI assistant, using Amazon Bedrock AgentCore and Model Context Protocol (MCP) to consolidate structured knowledge—service catalogs, API specifications, IT documentation, and data schemas—into a single governed source of truth (AWS Machine Learning Blog).
The architecture deployed in two steps. First, HEMA built HAL as a standalone chat assistant backed by a Strands agent on AgentCore runtime, connecting to Amazon Bedrock Knowledge Bases for semantic search and to internal APIs via an AgentCore Gateway (AWS Machine Learning Blog). Second, the team opened HAL to external MCP clients—including Kiro and Claude—through a second Entra ID-authenticated Gateway, allowing engineers to access the same knowledge from their IDE or preferred AI assistant without AWS credentials (AWS Machine Learning Blog). For commerce practitioners, this pattern demonstrates how to unify internal knowledge infrastructure once and distribute it across multiple tools via standardized protocols, eliminating per-tool custom integrations and reducing friction in engineering workflows.
HEMA plans to extend HAL from read-only knowledge layer to an action layer, enabling teams to execute operational tasks through the same interface (AWS Machine Learning Blog). The approach anchors security in Microsoft Entra ID with managed OAuth, avoiding the operational burden of distributing AWS credentials to end users.