The Connector Economy: Why the Next Platform War Is About MCP Hubs

In February 2026, Cloudflare demonstrated something that should worry every integration platform on the market. They took an API with 2,500 endpoints, compressed it into two MCP tools, and cut the token cost from 2 million to 1,000. Not a 2x improvement. A 2,000x improvement. The same month, France's data.gouv.fr shipped an MCP server for public government data. OpenAI, Anthropic, Google, and Microsoft are all standardizing on MCP as their agent integration layer. The ecosystem already has over 5,800 MCP servers live.

Zapier processes over 2 billion tasks per month. Make handles millions of scenarios daily. Workato powers enterprise automation for Fortune 500 companies. Together, these platforms built a $15 billion integration market by solving a real problem: business software doesn't talk to itself.

But they all share the same architecture. Trigger, action, repeat. When a new row appears in Google Sheets, create a contact in HubSpot. When a deal closes in Salesforce, send a Slack message. Each connection is a thin wire between two endpoints, and every wire must be hand-strung by a human who understands both tools.

This worked when companies used 20 apps. The average mid-market company now uses 130. And AI agents doubled their share of web traffic on Cloudflare's network in January 2026 alone. The tools aren't just multiplying. The consumers of those integrations are shifting from humans to machines.

The spaghetti problem

I learned this the hard way building integrations at Qonto. We needed to connect our banking product to accounting tools, expense platforms, and ERP systems across multiple European markets. Each integration was its own project. Each project required understanding the source API, the destination API, the data mapping between them, the error handling, and the edge cases that only surface at 3 AM on a Friday when nobody is watching.

We weren't unusual. Every SaaS company building in the fintech and B2B space faces the same grind. You ship the product, then spend a disproportionate amount of engineering time wiring it to the tools your customers already use.

Trigger-action platforms helped, but they didn't solve the core problem. They moved the complexity from code to configuration. Instead of writing integration code, teams built elaborate multi-step Zaps with conditional branches, filters, formatters, and retry logic. The spaghetti didn't disappear. It migrated from the codebase to the workflow editor, where it became someone else's problem.

At 50 integrations, you need a full-time person managing your Zapier account. At 200, you need a team. The tool that was supposed to eliminate busywork became busywork itself. I've talked to ops leaders at growth-stage startups who spend 15 to 20 hours a week maintaining their automation stack. That's a half-time job that didn't exist five years ago.

From recipes to intent

The next generation of integration doesn't work this way. MCP (Model Context Protocol) hubs replace trigger-action pairs with something fundamentally different: intent-driven orchestration.

Instead of building a workflow that says "when X happens in tool A, do Y in tool B," you describe what you want in plain language. "Sync all new invoices from QuickBooks to my CRM and flag any over €10,000 for review." The AI layer determines which connectors to use, how to map the fields, when to run the sync, and how to handle exceptions.

This is not a marginal improvement. It's an architectural shift.

Trigger-action platforms are protocol translators. They move data from point A to point B along a predefined path. MCP hubs are intelligent routers. They understand context, chain multiple tools in a single operation, and adapt when conditions change. If QuickBooks updates its invoice schema, a trigger-action workflow breaks. An MCP hub notices the change, adjusts the field mapping, and logs what it did.

The difference matters because it changes who can build integrations. Trigger-action workflows require someone who thinks in systems: inputs, outputs, conditionals, error states, retry policies. Intent-driven orchestration requires someone who can describe a business outcome. The second group is orders of magnitude larger than the first. A sales manager who has never seen an API can now orchestrate a five-tool workflow by describing what the result should look like.

Why connectors compound

Here's where the economics get interesting.

In a trigger-action platform, each new connector creates linear value. Connecting Stripe gives you Stripe workflows. Connecting Notion gives you Notion workflows. The value of connector number 200 is roughly independent of connectors 1 through 199. Each one opens a set of recipes, but those recipes don't interact with each other in meaningful ways.

In an MCP hub, each new connector creates combinatorial value. When the AI can access QuickBooks, Salesforce, and Slack simultaneously, it can execute operations that span all three in a single intent. "Check if any customer with overdue invoices in QuickBooks has an open deal in Salesforce, and if so, send their account manager a Slack summary." That operation touches three tools, requires contextual reasoning across all of them, and would take a dozen Zaps to replicate.

Add a fourth tool and you don't add one new capability. You add combinations with every existing connector.

This is Metcalfe's Law applied to software connectors. A hub with 100 connectors isn't twice as valuable as one with 50. It's potentially thousands of times more valuable, because the number of possible multi-tool operations scales combinatorially with the connector count.

This creates a flywheel that first-generation platforms never had. More connectors attract more users. More users generate more intent data, which improves the AI's ability to route and map between tools. Better routing makes the platform more reliable, which attracts more connector builders. The cycle accelerates.

The 500-connector threshold

There's a critical mass question: how many connectors does a hub need before the network effects become self-sustaining?

Based on patterns in adjacent platform markets, the answer is somewhere around 500. At that point, a hub covers roughly 80% of the tools any given company uses. Users stop asking "does it connect to X?" and start assuming it does. The mental model shifts from "let me check if this integration exists" to "let me just describe what I need."

That assumption is the unlock. Once users default to the hub as their integration layer, switching costs compound. The AI learns their data patterns, their field mappings, their business logic, their exception handling preferences. Every week of usage deepens the relationship. After six months, the hub knows that "customer" in your Salesforce maps to "client" in your invoicing tool, that your finance team prefers weekly batch syncs on Monday mornings, and that invoices from your German subsidiary need a different tax field mapping than the French ones. The hub becomes the connective tissue of the entire tech stack.

Think of this as the USB-C moment for enterprise software. For years, every device manufacturer shipped a proprietary connector. Mini-USB, micro-USB, Lightning, barrel jacks, magnetic adapters. Then a single standard emerged and the market consolidated around it. Not because USB-C was technically superior to every alternative on every dimension, but because universality creates its own gravity. The standard that covers 80% of use cases wins 100% of the market.

Data gravity and the hub position

The hub that reaches critical mass first captures something more valuable than subscription revenue: data gravity.

Every intent that flows through the hub generates metadata. Which tools do companies connect first? What operations do they run most frequently? Where do data mappings fail? Which field combinations indicate that a user is about to churn? This intelligence creates a moat that is nearly impossible to replicate through engineering alone. A competitor can rebuild your connector library in a year. They cannot replicate millions of learned intent-to-action mappings.

The hub also occupies a unique position to capture infrastructure-level economics. Just as AWS charges for compute, and Stripe charges for payment processing, the MCP hub charges for orchestration. The company that controls the routing layer between enterprise tools sits at a toll booth on every transaction that crosses tool boundaries.

This isn't rent-seeking. It's value creation. Orchestration is hard. Getting five tools to cooperate on a single business operation, handling authentication, rate limits, schema differences, and partial failures, is genuinely difficult work. Anyone who has built a multi-system integration from scratch knows: the first 80% takes a week, and the remaining 20% takes a month. The hub that compresses that timeline from weeks to seconds earns its margin.

The race that's already started

The incumbents see this shift coming. Zapier has been adding AI features to its workflow builder. Make is building natural language interfaces. Microsoft is positioning Copilot as an integration layer across its ecosystem. But retrofitting intelligence onto a trigger-action architecture is like adding a touchscreen to a flip phone. The underlying form factor constrains what's possible.

The real competition will be between AI-native hubs designed from day one around intent-driven orchestration. Platforms that treat connectors not as endpoints for a workflow builder, but as capabilities for an AI agent that reasons about business operations.

This race has three dimensions. First, connector breadth: who reaches 500 first and locks in the network effects. Second, intent intelligence: whose AI best translates business language into reliable multi-tool operations. Third, uptime: enterprise buyers need their integration layer to work every time, and many AI-native startups haven't yet proven they can deliver that consistency.

The connector economy is forming now. The integration market is restructuring around AI-native orchestration, and the timeline is three to five years. Given the speed at which enterprises are adopting AI tooling for operations, that estimate might be conservative.

At Well, we're building toward 500+ MCP connectors because we believe the hub that reaches critical mass first will define this category. Not through features, not through pricing, but through the gravitational pull of a universal connector network that gets smarter with every intent it processes.

Maxime Champoux is the co-founder and CEO of Well.