Ship an MCP or become a database.

Greenhouse shipped an MCP server this year, rolling it out to customers through June, and the part I found beautiful was the demo itself. A recruiter opens Claude, asks for the state of three open roles, gets back a pipeline report with candidate summaries and a recommended next step, and clears a batch of rejections, without loading greenhouse.com even once. The software did the work. The recruiter did not even go to their dashboard or type in the domain name.

MCP is the Model Context Protocol, the open thing Anthropic shipped in late 2024 and then handed to the wider ecosystem, and the cleanest way to hold it in your head is Anthropic's own line that it works like a USB-C port for AI. A product exposes a set of governed tools, the agent calls them on the user's behalf, and the data flows straight into the conversation the user was already having. Greenhouse is not special here. It is the cleanest public version of a move I have been watching arrive quietly across every product that lives on customer data, where the customer keeps the outcome and quietly stops opening the app that produced it.

For about fifteen years the SaaS dashboard was the product.

The screen was where you fought for the user, where the design team earned its budget, where the demo lived and the moat was supposed to sit. The whole discipline of B2B product design grew up around making that screen good enough that a buyer would pick you and a user would stay. And the screen is now becoming optional, because for a growing slice of your usage the user enters through Claude and never arrives at the thing you built.

I spend my days on the integration surface of a B2B platform and on the agentic layer we built over it, so I have watched this arrive from the supply side rather than read about it. The thing that took me a while to accept is how much of the dashboard quietly stops mattering once the user lives in the agent. Users never browse your navigation, so the careful information architecture you argued about for a quarter goes invisible, and a chat reply carries no logo and no brand colour, so your brand becomes whatever tone the model happens to use when it mentions you, if it mentions you by name at all. The funnel you instrumented turns into prompt logs you do not own, and the upgrade banner simply has nowhere left to render. Even the switching cost you spent years building, the muscle memory of where every button lives, drops to the price of editing one line in a tool config.

The whole threat compresses into one sentence: when the user lives inside someone else's agent, a product without a good MCP is just a database the agent hasn't bothered to connect to yet.

The framing that makes this legible to a CFO is Ben Thompson's Aggregation Theory, which argues that whoever owns the customer relationship commoditises everyone upstream of it. The agent is becoming the aggregator of software itself. If your only relationship with the user now runs through Claude or ChatGPT, you are a supplier, and suppliers get compared on price, latency, and availability rather than chosen out of loyalty to a screen nobody looks at anymore. The way you avoid that fate is to make your MCP the best one in your category, good enough that the agent reaches for you first and represents you accurately when it does.

None of this is the first time the interface decoupled from the product. Headless commerce and the API-first wave already separated the thing that holds the data from the thing that renders it, and a generation of companies learned that the API could be the product. What is new is that the renderer is now a language model, the user is talking to it in plain English, and the model decides moment to moment which of your tools to call and how to describe what came back. The decoupling went one layer deeper, and it took the user's attention with it.

The part Greenhouse got right, and the part most teams will get wrong, is that the MCP is governed. Every call goes through a defined tool, tied to the permissions the customer already has, with an audit trail behind it, which is the only reason a security or legal team will let an agent near production data at all. So shipping a good MCP takes choosing the dozen jobs that actually matter and naming them so a model selects the right one under pressure, which is a very different exercise from spending a weekend wrapping your endpoints. Then it is shaping each output so the model narrates it without inventing a number, and putting permissions and logging around the whole thing that your compliance team can actually read. That is product work and design work, not a plumbing task you hand to whoever is free.

And the dashboard does not die, which is the nuance the doom version of this argument skips. The interface splits in two. The destination UI still owns first run, trust, exploration, and the decision that carries real stakes, the kind a user wants to see with their own eyes before they commit, and a recruiter making a final offer or a marketer approving a send that could move a million dollars is going to open the real screen. The MCP owns the recurring execution, the report you run every Tuesday morning, the bulk operation, the question the user would rather just ask out loud. You now have two front doors, and the mistake is assuming the dashboard is safe because it is pretty, when half its traffic is about to move next door.

So the question I have started asking on our own roadmap is what the product sounds like when an agent reads it out to a marketer who will close the laptop having used us all afternoon and never once seen our name. When I first asked it in a planning meeting, the honest answer was that we sounded like a pile of undocumented endpoints a model would give up on halfway through. We are about nine months from having a better answer, assuming nothing gets deprioritised in October, which it will.