Alex, After The Great Library at Alexandria

TL;DR: One front door over every place our knowledge lives, so nobody — human or agent — has to remember which system holds the answer.

The knowledge a small company runs on is never in one place. Ours lives in Microsoft 365, in meeting transcripts, in project documents, and in a steadily growing pile of things that seemed sensible to put somewhere at the time.

Which is fine when you’re the only one looking, and stops being fine the moment you have AI agents that need to answer questions. An agent that has to know which system holds the answer before it can find the answer isn’t a research assistant. It’s a very expensive filing clerk with a memory problem.

So we built Alex.

The idea

Alex is the front door to everything Cairo knows.

You ask Alex a question. Alex works out where the answer lives, retrieves the relevant material, and hands it back in one consistent shape. You never query an individual backend. You ask Alex.

The name is short for Alexandria — the greatest knowledge repository of the ancient world, and, fittingly, in Egypt. This is Cairo’s library. We take the theme seriously around here.

The shape

        You / Claude / the pipeline
                    │
                 ┌──▼───┐
                 │ ALEX │   the router — the thing you ask
                 └──┬───┘
                    │  fans the question out, merges results
      ┌─────────────┼─────────────┐
      ▼             ▼             ▼
  ┌────────┐  ┌───────────┐  ┌─────────┐
  │  M365  │  │transcripts│  │ future  │
  └────────┘  └───────────┘  └─────────┘
   SharePoint    Fireflies
   OneDrive
   OneNote

Every source speaks the same neutral contract — a “knowledge extract” shape that says nothing about where it came from. That’s the design decision the whole thing rests on.

Because sources are wildly different. One is a document store with permissions and folders. Another is meeting transcripts with speakers and timestamps. If the router had to understand each one’s native shape, adding a source would mean changing the router, and the router would slowly accumulate a little piece of every system it touches until it was the most fragile thing we owned.

Instead: each source is an independent connector that knows how to translate itself into the common shape. The router only speaks the common shape. Adding a source is writing a connector, not editing the core.

What the router earns

Merging results without co-mingling stores.

That distinction matters. Alex can fan a question out to several sources and give you one coherent answer — but the sources stay separate underneath. Nothing gets copied into a central index that then has to be kept fresh, permissions and all. Each source remains the authority on its own content.

Which is what makes the security story tractable. If everything had been dumped into one big index, we’d have inherited the union of every system’s access rules and would be maintaining that union forever, correctly, or leaking. Keeping sources independent means each one keeps enforcing its own rules, and the router is a question-asker rather than a second copy of everything.

Why it’s an MCP server

Alex runs as an MCP front door, which means anything that speaks MCP can use it — my agents, the pipeline, a chat session, a future thing I haven’t thought of.

That’s the layering I keep coming back to: live data goes through MCP, timeless knowledge goes in a document. Alex is live by definition. What’s in a meeting transcript from Tuesday is not something you write down once.

The bit I’d tell someone else

If you’re wiring AI into a real organization, the router is the piece worth building early and the piece people usually build last.

The instinct is to connect the agent directly to whichever system you need first, and then to the second one, and then the third — and each connection is easy, so it never feels like the wrong call. Six connections later, every agent knows about every backend, changing a source means changing everything that touches it, and the coupling is load-bearing.

One layer in the middle, one neutral contract. Everything upstream gets simpler, and adding the seventh source is the same amount of work as adding the second.

Named after a library, built like a switchboard.

The projects, experience and opinions here are mine. AI helped me turn my notes and build records into this piece and polished it for Cairoglyphics.ai.