Let me ground that in three things:
The Operating System (OS) of tomorrow.. which is really the Operating System of today.
The kinds of companies that have a connector, and how they feel about it.
Why one of those kinds is built, from the ground up, for the world that’s coming.
The Operating System
I'll use "MCP" from here on. It means connector. Apologies if you’re already fluent. Skip the next bit.
For a long time, businesses had a REST API. It meant a developer somewhere else, inside some other company, could reach in and pull data out in an orderly way. What an MCP does is wrap that same idea in a layer the AI can read on its own. It tells the model, in real time, what each tool is and what it does, so the model can go and query a company’s data, look at what comes back, decide it needs more, query again, find the answer, and carry it back into the chat with you. It’s not a handcrafted workflow but progressively discovered by an LLM.
The times are truly “a changing”. I had a moment that crystalised the future of work. Inside Claude, in a single swoop, I queried my financial data through Seapoint’s MCP, queried our legal data through SeedLegals’ MCP, and queried my Granola’s MCP for meeting transcripts. Resultantly I answered a complicated, situational, personal, domain knowledge required, query, in one go. It was unprecedented intelligence precisely because it’s personal.
That’s what I mean by the operating system of tomorrow being, really, the operating system of today. Because the future is already here, adoption is just not evenly distributed. All the data that used to be gatekept inside enterprise software, once locked behind a login and a dashboard I had to go and stand inside, is now queryable through a window. It’s already happening. And once you’ve felt it, it’s overwhelmingly obvious how we’re going to work.
Which makes you ask a slightly frightening follow-up: in that world, how much software do you actually need? Definitely not all of it, if that’s helpful. Here’s how I split the software into piles:
Software, It:
Displays.
Lets you make.
Makes for you.
Essentially only those categories across the board.
The dividing line is data. Pile one creates no new data of its own. It’s a nicer window onto data that lives somewhere else. Pile two lets you generate the data by hand. Pile three generates the data itself.
Names, so it isn’t abstract:
DocuSign = pile one. While a signature is technically new data, it’s nominal. The substance was created elsewhere.
Figma = pile two. You make a design in it you genuinely couldn’t have made without it. The data is yours, and it wouldn’t exist otherwise.
Granola = pile three. It takes the most analogue thing in your working life, people talking in a room, and hands you data you could not have had before.
Hold that gradient in your head. As the OS revolution eats into “software,” these three piles are not equally at risk or equally positioned.
The types of companies that have a connector
Now walk each one up to the door and ask it to ship a connector. See if they flinch at all.
Pile one is scared, and it maybe really should be. A connector means a user can pull everything out and see it in the cold light of an agent, and notice exactly how little was ever there. We don’t really do anything, and people might notice. The moment your data is portable, the question “why do I open this app?” answers itself, and the answer is “yeah so.. maybe don’t.” It can be defensible, if they commercialise an API, and the data is totally unachievable elsewhere.
Pile two is not totally sure. It holds something real so there’s a genuine pull to connect, because you’ll want your own work queryable and actionable from wherever you happen to be. The user experiencing no friction is important, within a pile of companies who are also structurally competing the heaviest on their user experience. But the interface is also their product. If your agent can read and act on the Figma file without ever opening Figma, the reason to open Figma gets thinner. If coding the design is quicker and-or better, the Figma MCP is overlooked in the session. The strategic companies within this pile are planning their transition into pile three, so that they can be a default for agents.
Finally, pile three has nothing to be nervous about. It is structurally incentivised, wholly, to be everywhere and anywhere all at once 🥯. Because in the future’s/today’s operating system, apps are data sources. And pile three’s entire reason to exist is generating data that is yours and that didn’t exist before itself. That is the exact thing an agent goes looking for. Claude queries my transcripts, my messages, my finances. Claude does not need to query my weather app because there’s nothing in there that’s mine, or anything that only lives there. Pile three is the thing holding the one copy of something that is actually required.
So, the game theory here, truly:
In a world where the agent is the surface you live on, your value to a person is entirely downstream of whether their agent can reach you. Connect, and you get queried. You become infrastructure, you become depended on a hundred times a day. Don’t connect, and the agent just routes around you: it finds the data elsewhere, or it finds a competitor who did, and the user experiences you as friction. For anyone holding data that’s genuinely yours and genuinely unique, opening the door is the dominant, and arguably only, strategy.
The question becomes, are we making valuable data? Pile 3 companies. It’s totally possible to. With LLM’s, every generation is new data for the user. The follow-up is whether it’s defensible, and whether it’s better you did it the way you do it, than if I did it in Claude. But usually the best answer here is structurally easy; meeting note apps are the best example. Otter.ai, Granola.ai, though simple, create a meeting note and transcript that you could not have before.
Does this explain the hoarders?
Hoarders. Said with disdain. The companies that break the model. LinkedIn. You can do things in them, they show you things, they generate valuable data and yet they are absolutely, deliberately bolted shut from an integration point of view. I can tell you that with feeling, because at Cloop we would love to integrate, yet they go out of their way to be unlovable 👉👈.
There’s an entire grey-area SaaS economy built on nothing but reverse-engineering automations for it. Tools quietly breaking LinkedIn’s terms of service, tolerated precisely because that scraping feeds the funnel into Sales Navigator, LinkedIn’s own paid data product. Doing it properly is almost certainly somewhere on their roadmap, but they don’t want to lose a minute of on-platform traffic before then, so they keep the lawsuits and the bans holstered for later. You can ask Proxycurl, which LinkedIn litigated clean out of existence.
I imagine LinkedIn only lets you export your data at all because GDPR compels them, and that export is the single reason a popular tool like Happenstance can exist. And they were so snivelling and sore about doing it that they made you wait up to seventy-two hours to get hold of your own connection data. Let’s be honest with each other: it is ready instantly. It is sitting right there. It’s just three extra days for you to forget you ever wanted it. Forget Happenstance. Forget that you were about to go somewhere else.
LinkedIn is a pile-three company. It holds some of the richest data about you, should you use it, that exists anywhere. But it’s scared to open the floodgates. To be a fair narrator, maybe there’s fear around their users getting spammed should they open up an MCP. But then justify the grey-market of SaaS’s being allowed to exist, that importantly, do spam LinkedIn’s users.
Google worked it out twenty years ago. They half-jokingly called it the Data Liberation Front, whose entire job was to make it easy for you to leave. Google now has the best integrations across the board. Their logic was that they want you to stay because you prefer them, “not because you feel locked in.”
It’s not about LinkedIn
LinkedIn is just my example of a company standing at this fork.
The fork is this. For about fifteen years, you won by being a destination, you owned attention, you were the place people went, and you measured everything by whether someone opened your app. Daily active users. That whole game is ending. In this OS world the game is not to be necessarily opened, rather than it is to be read from and written to by an agent. By being the source an agent trusts and keeps coming back to.
It’s who floats.
If you were only ever a display layer, without defensible data, be careful. It’s like how the ChatGPT query-response experience took over any previous search engine experience. If your game is making sure people come back, then having defensible data and shipping an MCP might be really valuable. If you’re a genuine source. If the data only exists because you exist, then being read from is fantastic. An agent cannot replace your source and the distribution you earned to get it. Granola doesn’t lose when Claude reads its transcripts. It wins because now it’s the thing being read a hundred times a day, instead of the thing I open once a week, and its dependency in my “AI stack” is what keeps me paying.
So, my informal recommendation is to ship an MCP. Specifically before the grey market ships a worse one in your name. Make your data trivially reachable, deliberately, cleanly, and first.
It’s the same for any type of company. If you’re a displayer, go and find the one sliver of data that is genuinely, only yours and become the supplier of that. If you’re a hoarder sitting on a goldmine like LinkedIn’s, open it before someone opens it for you; be the clean, first-party supplier instead of the thing a hundred grey-market scrapers spend their days impersonating. People would pay a lot for LinkedIn MCP unfortunately.
Stop displaying, stop hoarding, start supplying. Start building now, with purpose, into the world that may only care about, and reach for, sources. And think forward around positioning and strategy. You are not a piece of software with some data behind it. You are a data source that, for now, still happens to have a UI, and the UI gets more and more optional. The companies that make it through this transition will be the ones that worked that out early.
At Cloop, we’re wholly betting that the future isn’t another app you open but a layer that reads across everything you’ve already got. Which is also why, right now, we’re rolling our own MCP server out to our cohort of alpha users. We’re becoming a source.
To float in the world we’re running into, you have to be reachable.
