The MCP Boom: How a Protocol Went From 2 Tools to 228 in Under a Year

Something shifted this year in how software talks to AI, and our catalog caught the exact shape of it. In late 2025 we listed two tools that shipped a Model Context Protocol (MCP) server. As of our August 2026 snapshot we list 228. That is a 114x jump inside a single catalog of 9,959 published tools, and it happened without MCP touching most of the software market at all. The protocol did not spread evenly. It went straight to the places already wiring machines to models: developer tools, AI assistants, API layers, and the agent stacks being built on top of them.

The most striking number is not the growth, it is the price. Across the 228 MCP tools, roughly 85% offer a free tier and 37% are fully free, against 14% for the catalog as a whole. That makes an MCP tool about 2.6 times more likely to be fully free than the average product we track. Read together, those two facts tell one story: MCP is being adopted the way protocols get adopted, by giving the connective piece away to win the integration rather than by selling it. If you want the mechanics of how an MCP server differs from a plain REST API, our companion piece MCP vs API covers the wiring. This report is the market map: who is shipping MCP, how fast, at what price, and where it has not landed yet.

How we counted this

This report is a census, not a survey. We looked at every tool in our published catalog (9,959 products across 401 categories as of the August 2026 snapshot) and flagged the ones that ship or document an MCP server. Each of those 228 tools already carries the two attributes this analysis depends on: at least one editorial category, and a pricing bucket (free, freemium, or paid). Nothing here is modeled or extrapolated. The counts are direct reads from records we maintain for every listing.

Two caveats keep the numbers honest. First, catalog counts are a lower bound. Not every MCP server on the internet is a listed product, and plenty of MCP endpoints ship inside larger platforms, open-source repositories, or internal tooling that never becomes a discrete listing. When we say 228, we mean 228 distinct products in our catalog that expose MCP, which is a floor on the real population, not a ceiling. Second, our categories are editorial and a single tool can carry more than one. A database platform with an agent-facing server can sit in both Data and Databases and API Tools. That overlap is deliberate (it reflects how the tools actually position themselves), and it means the per-category counts add up to more than 228. We call that out again at the data table, where the shares sum above 100% for the same reason.

Pricing is bucketed the way it is everywhere in our catalog. Free means no paid tier exists for the core product. Freemium means a usable free tier plus paid upgrades. Paid means you pay to use it at all. For the free-tier comparisons below, we treat free and freemium together as tools a developer can start using without a purchase order.

The growth curve: from a rounding error to 132 in a quarter

MCP's arc in our catalog is not a smooth ramp. It is a flat line, a jump, a pause, and then a wave.

The quarterly additions run 2, then 76, then 18, then 132. Cumulatively that is 2 tools, then 78, then 96, then 228. Two things stand out. The first is the standing start: for most of late 2025 MCP was a rounding error, two listings against a catalog in the thousands. The protocol existed, but shipping a server was still an early-adopter move.

The second is the shape of the take-off. Q1 2026 delivered the first real wave (76 new tools) as the spec moved from novelty to expectation among developer-facing products. Then Q2 cooled sharply to 18. A dip like that is easy to over-read, but the simplest explanation usually holds: the fast movers had already shipped, and the next cohort needed time to build. Q3 2026 is where the story turns from interesting to structural. 132 new MCP tools in a single quarter is more than the entire preceding history of the protocol in our catalog. That is not a handful of pioneers anymore. That is the middle of the adoption curve arriving at once, which is exactly the pattern you see when a connective standard crosses from optional to table-stakes inside its home market.

Who builds MCP, and for whom

If growth tells you how fast, category concentration tells you who. And the answer is unambiguous: the people building MCP servers are building for machines and for the developers who wire them up.

The top of the chart is a clean statement of intent. Developer Tools (84), AI Assistants (69), and API Tools (56) are the three largest homes for MCP, followed closely by AI Agents (52) and AI and Automation (38). These are not consumer categories. They are the categories where the customer is a builder, an agent, or another piece of software. MCP's whole reason to exist is to give a model a structured, tool-calling way to reach an external system, so it makes sense that the first products to adopt it are the ones whose job is precisely that: exposing capabilities, moving data, and orchestrating calls.

The middle of the distribution reinforces the point rather than complicating it. Data and Databases (32), Automation (27), and Productivity (20) all sit at meaningful but smaller counts. These are places where an MCP server is genuinely useful (letting an assistant query a database or trigger an automation), but where MCP is a feature bolted onto an existing product rather than the product's reason for being. Workflow Automation (16) and Security (13) round out the top ten, which tells you the protocol has begun to reach adjacent operational tooling without yet saturating it.

To see the concentration as a landscape rather than a ranking, we plotted the same ten categories on a two-axis map.

A word on how to read this map, because it is easy to mistake for a scorecard. The x and y positions are derived, not published metrics. The horizontal axis (MCP tool density) is a relative placement based on how many MCP tools sit in each category, and the vertical axis (ecosystem momentum) is a relative placement based on each category's share of the recent additions driving the surge. We built both positions from the same underlying counts you see in the bar chart. They are there to show relationships (which categories cluster as leaders, which are still emerging), not to assign a precise numeric grade to any one category. Treat the map as a way to see the shape of the field, and the bar chart and data table as the exact numbers.

Read that way, the map has a clear upper-right cluster: Developer Tools, AI Assistants, and AI Agents anchor the leader corner, high on both density and momentum. API Tools and AI and Automation sit as challengers, real but not yet dominant. Data and Databases and Automation are emerging, and Productivity, Workflow Automation, and Security occupy the near corner as genuine but still niche adopters.

What the map makes visible is the absence. There is no marketing category on it, no HR category, no finance stack. That is not an oversight in the chart, it is the finding. MCP has not spread to marketing, HR, or finance tooling in any material way. Every one of the top ten homes for MCP is a developer or AI-infrastructure category. The protocol is still living entirely inside the layer that builds and connects AI systems, and has not yet crossed into the departmental software that the rest of a company actually uses day to day. That gap is the single biggest tell about where MCP is in its life: it has won the builders and has not yet been handed to the business.

The economics of a protocol

Pricing is where MCP stops looking like a normal software category and starts looking like a standard being seeded. The pricing mix for MCP tools is close to the inverse of the catalog average.

Start with the paid tier, because that is where the contrast is sharpest. Only 15% of MCP tools are paid-only, against 45% of the catalog. Flip that around and the pattern is unmistakable: 85% of MCP tools give you a way in for free (37% fully free plus 47% freemium), while the catalog splits roughly evenly, with 54% offering a free tier of any kind (14% free plus 40% freemium) and 45% paid. The freemium share is actually similar between the two groups (47% for MCP, 40% for the catalog). The gap is almost entirely in the fully free bucket: 37% for MCP versus 14% for everything else, which is the 2.6x free-likelihood difference stated up front.

Why would a whole class of tools give itself away at nearly triple the normal rate? Because for MCP, the server is rarely the product. It is the on-ramp to the product. An MCP server exists so an assistant or an agent can reach some underlying system, and the company shipping it almost always makes its money on that underlying system: the database, the API platform, the developer tool, the seat. Charging for the connector would tax the very adoption the connector is meant to unlock. The rational move is to make the MCP server free, get embedded in as many agent workflows as possible, and monetize the thing the agent is now reaching. That is a classic protocol land-grab, and the pricing data shows it happening in real time. MCP, right now, is an adoption line, not a revenue line.

The layer above: agents versus servers

It is tempting to read 228 and hear a big number. It helps to place it against the layer directly above it.

Our AI Agents category holds 773 tools. MCP servers number 228. The agents outnumber the servers more than three to one. That ratio is worth sitting with, because it maps the relationship between the two layers. Agents are the consumers. MCP servers are the connective tissue that lets those agents reach the outside world in a standard way. Today there are far more agents than there are standardized ways to plug them into anything.

Two readings follow, and both are probably true. The optimistic one: MCP has enormous headroom. If even a meaningful fraction of those 773 agents come to expect standardized tool access, the population of MCP servers has a long way to grow before it matches demand, and Q3's 132-tool quarter suggests that catch-up is underway. The cautionary one: a lot of agents are still integrating the old way, with bespoke API glue rather than a shared protocol, which is exactly what MCP is trying to replace. The gap between 773 and 228 is the size of the opportunity and the size of the work still to be done, in the same figure.

The MCP footprint by category

The table below lists the top ten categories by MCP tool count, each category's share of the 228-tool MCP set, and a short read on what that category represents. Because a single tool can carry more than one category, the shares sum to more than 100%; that overlap is the point, not an error.

The distribution has a long, developer-heavy head and a thin tail that stops well before it reaches business software. Developer Tools alone accounts for more than a third of the MCP set. The top four categories (Developer Tools, AI Assistants, API Tools, AI Agents) together describe nearly every MCP tool worth talking about. Everything below AI and Automation is real but secondary, and everything a marketing, HR, or finance team touches is simply not here yet.

What this means

Where this goes next

The pattern in our data is the same pattern most successful protocols follow: win the builders first, give the connective layer away, then wait for the standard to diffuse outward into the software everyone else uses. MCP has finished the first act. The 132-tool quarter, the 85% free-tier rate, and the near-total concentration in developer and AI-infrastructure categories all describe a protocol that has been adopted enthusiastically by the people closest to the model and not yet by anyone else. The interesting question for the next few snapshots is not whether MCP keeps growing inside its home turf, it clearly will, but whether it finally crosses the line into marketing, HR, and finance stacks. When that first wave shows up in our catalog, it will mean MCP has stopped being a builder's protocol and started becoming an everyone protocol. We will be counting when it does.