Early in an ecosystem, the observation is simple: nothing exists yet. The gap is everything.

As the ecosystem matures, tools start appearing. Not all at once, not uniformly, but in a recognizable pattern. The data layer gets built first — databases, APIs, connectors to existing systems. The knowledge layer comes next — models, embeddings, expert context that makes raw data useful. The workflow layer fills in last, because it requires the other two to exist before it can integrate them.

When you’re watching a nascent MCP ecosystem form in a specific vertical, you see this sequence play out in real time.

The first tools are data connectors. Someone wraps existing data APIs in MCP servers — property records, market comps, public databases. These are the cheapest tools to build because the data already exists. The value is just accessibility: data that used to require opening a portal can now be queried by an AI agent.

The second wave is knowledge infrastructure. Domain experts build MCP servers that encode their expertise — financial modeling conventions, analysis frameworks, market interpretation. This layer is harder to build because it requires deep domain knowledge, not just API wrappers. But it’s high-value: it lets an AI agent reason about data in ways that require professional judgment.

The third layer is the document workflow. Real work in most professional domains generates documents — contracts, reports, financial statements, structured data exports. These documents need to flow into the workflow. Extracting structured information from unstructured documents, normalizing formats, connecting document output to the rest of the stack — this is often the hardest layer to build and the last to be MCP-native.

When a stack is forming, the interesting question isn’t “is there anything yet?” It’s “which layer is still missing?”

The tools that fill the missing layer inherit everything the layers above them have built. If the data layer is already covered, the document workflow tool doesn’t have to solve the data problem — it just has to connect to what already exists. The stack does most of the work. You’re building the missing piece, not rebuilding the whole thing.

The gap in a maturing ecosystem is smaller and more specific than it looks from outside. But it’s also more defensible. You’re not competing with nothing — you’re building the tool that completes the stack. +++