What the $28.36 billion MCP market figure actually measures

$28.36 billion. That is the figure a market report puts on the Model Context Protocol, and it circulated on 14 September through GlobeNewswire. It is a big number for a protocol that was announced less than two years ago and that, technically, is little more than a specification for a model to call external tools through an explicit contract.

Before quoting that figure in a deck, it is worth asking what exactly it measures. Reports of this kind do not count MCP licences, because MCP is not licensed: it is an open specification. They count spending associated with an ecosystem, and the boundary of that ecosystem is drawn by whoever writes the report.

What tends to sit inside a number like that

In practice, a market of this type aggregates line items that already existed under another name: integration platforms, middleware, consulting services, observability tooling, API security and custom development. If a company pays a vendor to connect its CRM to an assistant, that contract counts. Whether the connector speaks MCP or speaks REST changes the invoice very little and the label a great deal.

This does not mean the report is wrong. It means the number is an attention label, not an operational data point. The press release is also the shop window for a paid report, so the headline is optimised to be quoted, which is exactly what is happening right now, this article included.

It is also worth looking at who publishes. GlobeNewswire is a press release distribution service: the firm that commissions the study pays to circulate it, and the outlet that picks it up does not verify the methodology. That does not invalidate the work, but it changes the weight it deserves. A sector report is judged on its methodology, its base year and its market definition, and none of those three things fits in a headline.

The signals you can actually check

There are indicators more useful than a projection, and they are observable without buying anything:

1. The standard is no longer single vendor. MCP was born at Anthropic, was published as an open specification, and adoption outside its own ecosystem is what turns a format into a standard.
2. MCP servers have left the developer tooling niche. On that same 14 September, an MCP connector for business payments was announced, which puts regulated systems on the other side of the protocol.
3. Compatible hosts have multiplied. Claude Code and Claude Desktop from Anthropic, plus third party clients, IDEs and in house agents built on the specification.

Any of those three facts says more about the health of the ecosystem than a multi year projection does.

What to do with this if you are building

For a technical team, the practical conclusion does not change with the size of the market. It changes with the cost of maintenance. An MCP server is easy to stand up in an afternoon and expensive to sustain for a year: specification versions, authentication, per tool permissions, usage limits, call logging, and the uncomfortable part, deciding which operations you will never expose.

The question we ask clients who arrive asking for 'an MCP' usually defuses half the project: which three concrete tasks do you want the agent to resolve without supervision? If the answer is a list of twenty, what they need is an API, not an MCP server. That conversation is more useful than any market projection, because it forces you to name the operations the agent can touch and the ones it cannot.

Our take

Market figures are mostly good for unlocking budget, and that is a legitimate use. We watch a different indicator, a less flashy one: how many of the integrations we build are still running six months later without anyone touching them. That number appears in no report and says rather more about whether a protocol has really settled.