Your Product Inside Their Assistant: The SaaS Funnel Loses Its First Three Steps
Nothing to install, no landing page, no empty state. What MCP changes about SaaS acquisition and activation, plus the five levers that cost a week.
A freelancer wants to know what Marc owes her this month.
First path. She searches for invoicing software, lands on a homepage, reads three bullet points, clicks "Start free trial", fills a form, verifies her address, arrives at an empty dashboard that invites her to import her clients. She closes the tab. Three days later a nudge email arrives that she never opens.
Second path. She pastes a URL into her assistant's connectors, approves one screen, asks her question, gets the amount. Forty seconds. She never saw your homepage, she does not know your name and she has used your product every day since.
For twenty years, SaaS acquisition has meant bringing someone to you. An assistant reverses the trip: your product goes to them, inside the tool they already have open. Your funnel loses its first three steps. What remains is shorter, harder to measure and harder to defend.
This article is for founders wondering whether to go, what to expect and where to start. It stays at the level of the decision and does not go down into the protocol. It is the conversation I have most often right now at the start of an AI-first web product engagement, before a line is written. For what AI changes about scoping a product, see the AI MVP playbook.
What changed, in two numbers
The first number says the shelf is free. I paginated the official MCP registry API1 on 18 September 2026: 33,350 unique servers, 9,812 of them published in the first eighteen days of the month. Three accounts alone account for 15% of them. To publish, the registry asks you to prove ownership of your namespace (GitHub, DNS or HTTP) and removes unwanted content only after the fact1. It does not sort the servers, which moves the entire difficulty onto being chosen.
The second number says installation is gone. Here, month by month, is the share of published servers, computed from the same API1, that require nothing to download, just a URL to paste:
| January 2026 | April | July | September |
|---|---|---|---|
| 26% | 48% | 57% | 66% |
In six months the ecosystem went from "clone this repo" to "paste this address". That is the material condition for everything else. There is nothing left to configure, so there is nothing left to justify a thirty-minute setup path, on your side or theirs.
Your funnel loses its first three steps
Put the two paths side by side, step by step.
| Step | Classic SaaS | Through an assistant |
|---|---|---|
| Discovery | Search, content, ads | The need is already stated |
| First visit | Homepage | None |
| Signup | Form, address verification | Consent screen |
| Empty state | "Import your data to get started" | None |
| Aha moment | A few days, via nudges | First answer |
| Habit | Notifications, emails | The tool is already open every day |
The empty state disappears. It is the number one activation killer in SaaS and it stops existing. An empty dashboard asks the user for effort before they have received anything. In a conversation the order flips: the person arrives with a question already formed, your product answers it, value comes before effort. Everything your product team built to bridge that gap, the demo data, the guided tours, the onboarding checklists, the nudge sequences, becomes moot.
Discovery is no longer yours. Your user did not compare three solutions on a feature-matrix page. Her assistant chose for her among the tools it had, reading their names and a one-line description. The positioning work does not disappear, it compresses into a single sentence read by a machine. I unpacked that selection mechanism in the article on loading tools on demand: past a few dozen available tools, clients stop showing them all to the model and hand it an index and a search instead. Claude Code ships this as a feature of its own, MCP tool search2.
The habit comes for free. SaaS retention means getting someone back into an application they tend to forget. Here, the surface is already open every morning for other reasons. The habit already exists, so the job is to be useful inside it.
Three questions to settle before writing a line
These three questions are not technical, and they decide the outcome.
Question 1: is your product a verb?
Chase a client, find a document, draft a quote, check stock: those gestures translate into tools an assistant can call. A dashboard, a workspace, a Kanban view: they do not. Those are surfaces, which an assistant cannot call.
The test is one exercise. Name three things your product does, in three words each, with a verb first. If it comes easily, you have your first three tools. If you struggle, your value lives in the interface itself and you will ship a server nobody calls.
That question hides a deeper one, worth asking even if you never build an MCP server. A product whose value lives entirely in its interface is a product whose value lives entirely in something AI can now reproduce in a few hours.
Question 2: what goes in front of the wall?
The consent screen is the new signup form. It converts worse, because it asks for more, sooner, with scarier wording: instead of an address and a password, it announces that a piece of software is about to reach your data. It is the exact place where your activation dies.
The only counter is to put value in front of it. What, in your product, can work without an account? A calculation, a simulation, a search across your public data, an answer to a reference question. You let the first calls through, the person sees that it works, then authentication arrives when they want to act on their own data. That is a product call, and it gets settled before the first line of code.
Letting the first calls through does not mean going without authentication. An Astrix study3 published in October 2025 of more than 5,200 open-source servers found that 53% reach the services they relay with long-lived static keys and only 8.5% use OAuth. You inherit the security reputation of that category. Today it is a real acquisition cost with the teams who decide.
Question 3: who pays when a machine is calling?
Per-seat pricing rests on a simple assumption: one human, one login, one subscription. An assistant calling your tools a hundred times a day for a single person breaks the link between usage and seat count. In the other direction, one licence can now serve a whole team if the configuration travels through the organisation.
You do not need to move everything to usage-based pricing tomorrow morning, but you do need to know, before opening this channel, what you are charging for: access, a volume of calls, or an outcome. Open the channel first and ask later, and you will discover your first serious users at the same time as the fact that they fit none of your plans.
Five levers that cost a week
These five levers are provided for by the protocol. They are the ones I push hardest in scoping, because the ratio of effort to activation is the best there is. The protocol and its layers are described in the article on MCP and A2A.
Lever 1: ship the menu with the kitchen
The specification4 lets a server provide ready-made starting points that the host displays to the user as commands to pick from. That is your getting-started guide, delivered through the channel itself, shown at the exact moment the person wonders what to do with the thing they just connected.
The official tutorial5 lists prompts among the three capabilities of a server, then concentrates on tools: the server it builds exposes two and no starting point. While everyone polishes their tools, the "here is what you can ask me" slot is easy to leave empty.
Lever 2: ask the question at the moment of the question
The specification also lets your server request information during execution6 rather than demanding it upfront. It is the difference between a twelve-field form at the door and a question asked when it becomes necessary.
In practice, every field you used to collect at signup becomes a question you will only ask the people who get far enough to need it. The rest of your form you can delete.
Lever 3: let every answer carry the next step
Everything your product returns is read by the assistant and rephrased to the user. A tool response that ends with a suggested next step will be relayed, in plain language, at the moment the person has just obtained a result.
It works like an onboarding email that arrives with no delay and that the person reads for sure. The guardrail: that slot has to be respected. A tool response that markets instead of helping teaches the model to prefer a competitor.
Lever 4: one user installs, the team inherits
A server's configuration lives in a file you can version alongside a project2, or in an organisation's managed configuration. One developer adds it to the repository, every teammate who pulls the code is asked to approve the server before first use. An admin can also provide it up front to every user in the organisation.
This distribution loop takes no extra work: your documentation simply has to show the shareable form of the configuration rather than the individual one.
Lever 5: your interface inside their conversation
An official extension to the protocol, announced7 on 26 January 2026 and developed by the MCP maintainers with OpenAI and the MCP-UI project, lets a server render its own interface directly inside the conversation. Your components, your colours, your buttons, in a sandboxed frame the host displays after the answer. The user clicks inside your product without leaving ChatGPT or Claude. The model stays aware of what they just did.
The announcement lists it as available in Claude on the web and in the desktop app, rolling out in ChatGPT and present in VS Code Insiders. Cursor added it8 in version 2.6. It wins back what the previous four levers make you give up: your brand, your ergonomics, visual recognition. A user who sees your interface remembers your name.
What you give up
This channel has a price: better to know it before you open.
The brand. Without lever 5, your product becomes a voice inside somebody else's application. The user from the opening gets her amount, she is satisfied, and she would be unable to name you. The day a competitor offers the same tool, there is no preference to overcome.
The measurement. No referrer, no tag, no journey to reconstruct. Your analytics will see nothing. If you want to know what this channel brings you, nobody will do it for you: you have to count calls per tool, identify the customers querying you and build the funnel yourself, from day one. Otherwise you will fly blind on a channel you cannot say works.
Per-seat pricing. See question 3. This channel will cost you a pricing rebuild, sooner than you expect.
Control over your visibility. Your tool is only shown to the model if it surfaces from a search you do not own, in an index you do not administer, against 33,349 other servers in the official registry. You can work on your names and your descriptions, you cannot buy the top slot.
The relationship. No address captured, no welcome sequence, no direct channel. You are a tool in a catalog and the person who connected you owes you nothing.
None of these is disqualifying. Each one invalidates part of what your team knows how to do, which is reason enough to discuss them before shipping rather than after.
What I would do this week
- Write the three verbs. Three gestures your product performs, three words each, verb first. If the exercise resists, stop there and keep the question for your next strategy review.
- Write the line. One sentence per tool, in your users' words rather than your internal vocabulary. It is the shortest version of your positioning, and it is the one the machine that picks you will read.
- Draw the wall. Decide what works without an account and what requires one. Write the rule down before anyone codes.
- Ship three starting points. The three questions your best customers ask in their first week, provided as commands ready to pick.
- Reread your tool responses. Each one should be able to end with a useful next step. Write them the way you would write a support reply, not an API payload.
- Publish the shareable configuration. In your documentation, lead with the form that versions alongside a project.
- Put the counter in. Calls per tool, per customer, per day, from the very first call. Without it, "is this working?" will still have no answer in six months.
- Settle the pricing. What you charge for on this channel, decided before you have users on it.
Seven of these eight are product and marketing work. Only one needs a developer. The patterns for that technical point are in the guide to designing an agentic module.
Closing
The channel is open, it is free and it is already crowded. What gets won there comes from three accurate verbs, a description line written to be found and a first answer that lands before the person has had to supply anything.
That is good news for teams who can articulate what their product does. It is bad news for the ones whose value mostly lived in the welcome screen.
Your user may never open your application. She can still become your best customer.
Sources
- MCP, about the registry, accessed 10/29/2026.
- Anthropic, Claude Code, connecting MCP servers, accessed 10/29/2026.
- Astrix Security, State of MCP Server Security 2025, 10/15/2025.
- MCP, specification, prompts, accessed 10/29/2026.
- MCP, build an MCP server, accessed 10/29/2026.
- MCP, specification, elicitation, accessed 10/29/2026.
- MCP, MCP Apps announcement, 01/26/2026.
- Cursor, forum, version 2.6 and MCP Apps announcement, 03/03/2026.