There are two strategies for building professional AI tools: own the whole stack, or own one layer and compose with the rest.

Owning the whole stack is the enterprise play. Build the data connectors, the knowledge layer, the document processing, the reporting — all proprietary, all integrated, all sold together. The value proposition is coherence: everything works together because everything is ours. The price is a procurement process, a contract, and a sales cycle.

The composable position is different. You own one layer — the one that’s missing, the one nobody else has built yet — and you integrate cleanly with everything above and below you. You’re not trying to replace the data layer. You plug into it. You’re not trying to replace the modeling layer. You feed it. You do one thing well and let the protocol handle the rest.

The composable position looks weaker on paper. You’re not a platform. You don’t have lock-in from owning the whole workflow. A user can theoretically replace you with a competing tool that does the same layer better.

But in practice, the composable position has a durability that platform plays don’t. When you own one layer cleanly, you become the obvious choice for that layer — not because users can’t leave, but because switching requires finding something that integrates as well. If you’ve built the tool that fits correctly into an emerging protocol ecosystem, the ecosystem itself creates switching friction. Other tools are built assuming you’re there.

The platform play is also fragile in a different way: it requires winning every layer to win at all. If a better document processor exists, the user has to give up the whole platform to use it. That’s not composability — that’s negotiating against a bundle.

MCP created the conditions for the composable position to work in professional domains. When every layer speaks the same protocol, a single-layer tool can be production-grade. It doesn’t have to solve the entire workflow to be genuinely useful. It just has to solve its layer correctly, expose it through the protocol, and trust the stack.

The tools that will last in this ecosystem aren’t the ones that own the most. They’re the ones that fit the best. +++