Build1 publisher2 min readPublished
HEMA wired its existing service catalog into Kiro and Claude as MCP tools
The AWS post co-written with HEMA says the structured half of its internal knowledge was already in good shape, and that the procedural half had little written documentation to fall back on.
The Engineer · Build desk
What happened
- HEMA, a Dutch retailer with more than 750 stores, built an internal assistant called HAL on Amazon Bedrock AgentCore and exposes its knowledge sources to AI clients through Model Context Protocol tools.
- The retailer already had a service catalog mapping people to teams, teams to services, and services to the APIs they expose and the business capabilities they support.
- Procedural questions, such as how to request access to an API or how to get a new group provisioned, had no single home in that catalog or anywhere else.
- Security for the assistant is anchored in Microsoft Entra ID, and the post says there are no AWS credentials on the client.
- HEMA says an answer that once meant navigating three or four portals, sometimes across an entire afternoon, now arrives in seconds inside the IDE or chat window.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- contradiction The post's headline diagnosis, that the knowledge existed but was hard to reach, does not cover the layer it later calls barely documented, so a team copying the pattern has to work out which of its own two problems it actually has.
- decision Whether this build is cheap depends on owning a maintained catalog first; a shop without one is buying a cataloguing project with an agent on top, and should price it that way.
- exposure A read-only tool surface already wired into three clients is the cheapest place to add write tools, and HEMA plans exactly that, so the next wrong answer at HEMA could provision something.
Expose each knowledge source once as an MCP tool, and any compatible client can call it. The alternative is a bespoke integration per source, rebuilt for every client application [9]. HEMA names three clients consuming the same tools: the HAL web chat, Kiro, and Claude [10]. Under the bespoke model that is three implementations per source against one under MCP, a three-to-one ratio at today's client count, and better with each client added [1].
The post describes Amazon Bedrock AgentCore as a platform to build, connect, and optimize agents at scale, with any framework [17]. That sentence is product-page copy.
The two goals the post sets out are both plumbing. The first belongs to HAL: consolidate fragmented knowledge into one governed source of truth. The second belongs to MCP: deliver that knowledge to people inside the tools they already use [18].
The post's thesis is stated flatly: "The problem was never that the knowledge didn't exist. It was that the knowledge was hard to reach." [6] For the structured layer, that holds. The post says that "For anything about what exists and who owns it, the answer was usually available, if you knew where to look." [5]
The procedural layer is a different problem with the same symptom. As HEMA grew and onboarded new engineers, the post says, the ask-the-person-next-to-you model stopped scaling and "there was little written documentation to fall back on" [8]. Indexing fixes knowledge that exists but is hard to reach. Knowledge nobody wrote down has to be written before any protocol can serve it. The post, co-written with HEMA's Mauro Rallo and Patrick van der Plas, attributes slow onboarding for new joiners, inconsistent answers depending on where someone looked, and constant context-switching to that gap [14][15].
The afternoon-to-seconds claim is a measurement of HEMA's portals and HEMA's catalog. Two conditions have to hold before the number means anything elsewhere. The catalog has to be current enough that returning its contents is useful, and the procedural rules have to exist as text a tool can retrieve. HEMA had years of work behind the first condition, including structured data from the product information management engine and the data-mesh tables imported and organized into the catalog [4].
What to watch
- Whether HEMA publishes what its write tools can do, and which Entra ID groups authorize them, once HAL becomes an action layer.
- Whether the MCP client list grows past the HAL web chat, Kiro and Claude, since the client count is what moves the integration ratio.
- Whether HEMA says where the procedural answers now live and who keeps them current.