The short version: shipping an MCP server has stopped being an engineering checkbox and started being a distribution decision. This week two very different companies — Crunchbase and Axonius — shipped MCP servers on the same day, July 21, and the reason is the same for both: their buyer increasingly asks the question inside Claude or ChatGPT, not inside their app. An MCP server is how you show up in that conversation. And the day after this publishes, the stateless 2026-07-28 spec finalizes, which makes running one cheap enough that "we'll get to it" stops being a defensible answer.
If you build software a founder or an analyst pays for, this is the channel question of the second half of 2026. Here's the read.
What shipped this week#
Two launches, one day, opposite corners of the market:
- Crunchbase launched an MCP that puts its private-market intelligence directly into LLM answers — the company cites 39 billion live signals, 790 data fields per company, and 15 million predictions. The pitch is that a dealmaker can ask "which accounts are most likely to raise in the next year?" inside their assistant and get Crunchbase's answer, without opening Crunchbase (Crunchbase via Yahoo Finance).
- Axonius launched an AI Agent and an MCP Server that pipe reconciled asset-security context into whatever AI tool a security team already uses — ask a plain-English question about your attack surface, get an answer grounded in Axonius's asset model (Axonius).
Different verticals, identical playbook: take the proprietary data you already own, and make it callable from the place the customer is already standing.
The one idea: a server is a channel, not a feature#
Here is the distinction worth internalizing. A normal integration lets people who already chose you connect a tool. An MCP server does something different — it makes your product a surface the user's agent can reach without your product being open. The buyer asks a question in Claude; your tool gets called; your answer lands in the reply. You won the query at the moment of intent, and you never competed for a browser tab.
That is the definition of a distribution channel: being present where demand forms. For a decade the channels were search, social, and app stores. In 2026 a new one opened — the inside of the assistant — and an MCP server is the on-ramp. Channels compound; features don't. That's why "should we ship an MCP server" is a GTM question wearing an engineering costume.
The buyer asks the question inside the assistant now. An MCP server is how you're in the room when they do.
Why July 28 removes the last excuse#
The reason this is urgent and not merely interesting: the finalized 2026-07-28 MCP revision makes the protocol stateless. No session handshake, no sticky Mcp-Session-Id header — any request can hit any instance behind a plain round-robin load balancer with no shared session store.
Translated to a founder's budget: an MCP server is now an ordinary stateless web service — the exact shape you already know how to deploy, scale, and pay for. The old objection ("it needs sticky sessions and a session store, that's real infra") is gone. What's left is a decision, not a cost.
Should you ship one?#
Not reflexively. The honest test is what your value actually is:
- If your value is a lookup or an action an agent should take on the user's behalf — a price, an inventory count, a database query, a booking, a piece of proprietary data — ship a server. That's the surface MCP rewards, and every week you don't, a competitor's tool is the one that gets called.
- If your value is an experience a human needs to sit inside — a canvas, a document, a live workspace — an MCP server is a side door, not the front door. Build it, but later, and don't expect it to carry the product.
And whichever you ship, instrument it like a channel. Log which tools get called, from which clients, for which queries, and whether those calls convert. An MCP server you announce and forget is a press release. An MCP server you measure is a funnel — and the teams that win this will A/B their tool descriptions the way they A/B landing-page copy. (If you're weighing the raw economics of agent tooling before you commit, our read on where 2026 agent spend is actually going is the companion piece.)
The founder action, in one line#
Decide this week whether your product is a call or a place. If it's a call, the stateless spec that locks July 28 just made shipping the channel cheap — so ship it, and put analytics on it before you put a blog post on it.
