Your Next Million Users Will Be AI Agents. MCP Servers Are the New App Store.

Something structural is happening in software distribution and most people building products have not noticed yet.

MCP, the Model Context Protocol, went from a quiet Anthropic announcement in November 2024 to 97 million monthly SDK downloads across Python and TypeScript. There are over 10,000 active MCP servers. Agent skill marketplaces have published 350,000 skills in roughly two months. And the adopter list reads like a roll call of the most important infrastructure companies in tech: Stripe, Shopify, Google, Datadog, Cloudflare, GitHub, Asana, Zapier.

Aakash Gupta, one of the most-read product management writers on Substack, published a detailed analysis this week arguing that MCP servers are the fifth major software distribution channel, after retail, web, mobile, and AI discovery. The pattern he describes is familiar: every time the distribution interface shifts, the companies that build for the new channel first capture outsized market share.

The companies that are building for agents right now will own the next decade of software distribution. And there is an enormous amount of infrastructure, tooling, and services to build around this shift.

What MCP actually is

MCP is an open protocol that lets AI agents connect to external tools, databases, and services through a standardized interface. Think of it as USB-C for AI. Before MCP, every AI integration was custom. If you wanted Claude to access your Salesforce data, you wrote a bespoke integration. If you wanted GPT to query your database, you built a custom function.

MCP standardized this. An MCP server exposes tools and data sources through a consistent interface. Any AI agent that speaks MCP can connect to any MCP server. One protocol, universal compatibility.

Anthropic donated MCP to the Linux Foundation's Agentic AI Foundation in December 2025. OpenAI and Block joined as co-founders. Google, Microsoft, AWS, Cloudflare, and Bloomberg signed on as supporting members. When every major AI company co-governs a protocol, it is no longer a bet. It is the standard.

The agent skills explosion

The speed of the ecosystem forming around MCP is unprecedented. npm took a decade to reach 350,000 packages. The agent skills ecosystem did it in two months.

Three marketplaces have emerged. SkillsMP is the volume leader with over 351,000 skills, functioning like the npm registry for AI agent knowledge. Skills.sh is Vercel's entry with 83,000 skills and 8 million installs, focused on cross-agent compatibility across 18 different AI agents. ClawHub is the curated alternative with about 3,200 skills and 1.5 million downloads.

The comparison to the App Store's early days is apt. In 2008, the developers who figured out mobile distribution first built the biggest app businesses. The same dynamic is playing out now with agent distribution.

Why this creates business opportunities

Gartner projects that 40% of enterprise applications will embed task-specific AI agents by end of 2026, up from less than 5% in 2025. That is one of the steepest adoption curves in enterprise history.

But here is the gap: most SaaS companies have not shipped an MCP server yet. Most vertical software has zero agent accessibility. And the tooling around MCP development, testing, monitoring, and discovery is almost entirely unbuilt.

That gap is where businesses get built.

Vertical MCP servers for specific industries

The horizontal MCP servers exist. Stripe has one for payments. Shopify has one for commerce. Datadog just launched one for observability. What does not exist yet are MCP servers for the thousands of vertical software tools that small businesses use every day.

A dental practice management system that lets AI agents schedule appointments, check insurance eligibility, and send reminders. A real estate CRM that lets agents pull comparable sales, generate listing descriptions, and manage showings. A restaurant POS system that lets agents analyze sales data, adjust menu pricing, and manage inventory.

Each of these is a product. Build an MCP server that sits on top of the existing software, connects it to AI agents, and charges the business $99 to $499 per month for agent accessibility.

The moat is domain knowledge. Anyone can read the MCP protocol docs. Very few people understand the specific data structures, workflows, and compliance requirements of dental practice management or restaurant operations or property management. That domain expertise becomes your competitive advantage.

MCP integration consulting

Most SaaS companies know they need to build for agents. They read the Karpathy tweet. They see Stripe and Shopify shipping MCP servers. But their engineering team has never built one, and they do not know where to start.

An engagement looks like this: audit the company's existing API surface, identify the highest-value tools and data sources to expose through MCP, build and deploy the MCP server, write the agent-facing documentation, and test it across major AI agents like Claude, GPT, and Codex.

Price it at $5,000 for a basic integration and $15,000 to $25,000 for a comprehensive build that includes custom tool definitions, authentication flows, and multi-agent testing.

The sales pitch writes itself. Andrej Karpathy said "Build. For. Agents." Gartner says 40% of enterprise apps will have agents by year end. Your competitors just shipped their MCP server. Can your engineering team build one this quarter, or do you want us to do it in two weeks?

At three engagements per month at $15,000 each, a solo consultant generates $45,000 in monthly revenue. Add a second consultant and you are approaching $1 million annually.

MCP testing and monitoring tools

Developers building MCP servers need the same tooling that API developers have had for years. Request inspection. Response validation. Performance monitoring. Error tracking. Load testing.

Think Postman, but for MCP. A tool where you can browse an MCP server's capabilities, send test requests, inspect responses, validate schemas, and run automated test suites. Layer on monitoring: uptime tracking, response time percentiles, error rates, usage analytics.

The API testing market is worth $2.6 billion and growing. The MCP testing market is at zero. The first credible tool in this space has a real shot at becoming the default.

Price it at $29 per month for individual developers, $99 for teams, $299 for organizations. If 10,000 developers working on MCP servers each pay $29 per month, that is $3.5 million in annual recurring revenue from the individual tier alone.

A curated MCP server directory

SkillsMP has 350,000 skills but almost no quality signal beyond GitHub stars. Skills.sh tracks installs but does not review or compare. There is no "G2 for MCP servers" where a product manager can evaluate which MCP servers to integrate and which to avoid.

Build the directory. Categorize MCP servers by function and industry. Write detailed reviews. Create comparison pages ("Stripe MCP vs Square MCP" or "Best MCP servers for e-commerce"). Add user ratings and integration guides.

Monetize through affiliate links when MCP servers have paid tiers, sponsored listings for companies launching new servers, and a premium tier for detailed analytics and integration support. The comparison pages are an SEO machine, ranking for every "[product] MCP server" query that developers and PMs are searching.

Premium agent skills as a product

Most published agent skills are free and open source. That creates an opportunity for premium skills with significantly higher quality, reliability, and support.

A premium "enterprise sales agent" skill that knows how to use Salesforce, HubSpot, Outreach, and LinkedIn Navigator together. A premium "full-stack deployment" skill that handles Vercel, AWS, and database migrations as a coordinated workflow. A premium "compliance and audit" skill for regulated industries.

Sell them directly at $49 to $199 per skill, or through a subscription at $29 per month for access to the full library. The value proposition: free skills give you 70% of the way there. Premium skills handle the edge cases, error recovery, and multi-tool orchestration that make the difference between a demo and a production workflow.

The Shopify signal

On March 3, Shopify and Google jointly announced the Universal Commerce Protocol, a new open standard built on MCP that lets AI agents complete real purchases inside conversations. Shopify rolled native MCP support into every store on the platform.

This is not an integration partnership. This is the largest e-commerce platform in the world saying: the next generation of commerce transactions will happen through AI agents, not through web browsers.

When Shopify makes that bet, every e-commerce tool, every marketing platform, and every logistics service needs to be agent-accessible. The companies that build MCP servers for the Shopify ecosystem will capture the integration layer between AI agents and commerce.

Getting started

Building an MCP server is not complex. The protocol is well-documented. The SDK is available in Python and TypeScript. A basic MCP server that exposes three to five tools from an existing API can be built in a weekend.

The harder part is knowing which tools to expose, how to structure them for agent consumption, and how to handle authentication and rate limiting in an agent-first way. That is where domain expertise and design taste matter.

Start with your own product or a product you know deeply. Build the MCP server. Publish it. See how agents interact with it. The learning from building one is the foundation for everything else, whether you go the consulting route, the product route, or the marketplace route.

What to take from this

The distribution channel is shifting. MCP has 97 million monthly SDK downloads and backing from every major AI company. The agent skills ecosystem grew faster than npm. Shopify is rebuilding commerce around agent protocols. And the infrastructure layer, the tooling, consulting, vertical servers, and marketplaces that this ecosystem needs, is almost entirely unbuilt.

Pick an approach. Build this week.