Skip to content
AI

Agents are users now

Andrés Max Andrés Max
· May 13, 2026

An agent interface is any way an AI agent can use your product without a human clicking through your UI. That includes MCP servers, command-line tools, and clean public APIs.

Here’s the part most founders miss. When you ship one of these, your product becomes something an AI can recommend, invoke, and use on a buyer’s behalf. No homepage visit needed. No ad spend. The agent is now in the room with your customer, and it picks the tool that’s reachable. That’s a brand new distribution channel, and it’s mostly empty.

Why this changed in the last 12 months

Anthropic shipped the Model Context Protocol (MCP) in November 2024. In the year since, MCP went from a niche spec to the default way AI agents reach external tools. Linear, Notion, Slack, GitHub, Stripe, Figma, and dozens of others now ship official MCP servers. Mercury, the startup-focused bank, shipped a public CLI in 2025 so founders can run their entire banking workflow from a terminal. The thing that used to take six clicks and a 2FA prompt is now one command.

The pattern is clear. Every serious product is becoming reachable from outside the screen. Sometimes by humans through a CLI. More often by agents through MCP or an API.

What is an agent interface

An agent interface is the surface of your product that an AI can call without a browser. The three common shapes are:

  1. MCP server. A standardized protocol introduced by Anthropic that lets any AI client (Claude, Cursor, ChatGPT Desktop, others) discover and invoke your product’s tools.
  2. CLI. A command-line interface that wraps your core actions. Humans use it. Agents use it more.
  3. Public API. A clean, documented, agent-friendly API with stable endpoints and clear auth.

The interface layer used to be one thing: your web app or your mobile app. Now it’s plural. Most successful products built in 2026 ship at least two interfaces from day one.

Why agents are users now

Agents are now inside the buying loop. Your prospect is asking Claude or Cursor or ChatGPT to help them do something, and the agent reaches for the tool that’s available. When your product is reachable, you get included in workflows you never marketed to, by users who never saw your landing page.

Three concrete examples of how this plays out today:

  • A founder asks Claude to file an expense report. Claude calls the Mercury CLI directly. Mercury just got used inside a workflow Mercury didn’t have to advertise.
  • A PM asks ChatGPT to summarize this week’s tickets. ChatGPT calls the Linear MCP server. Linear just earned a daily touchpoint with a user who might never have logged in.
  • A developer asks Cursor to update a Stripe price. Cursor hits the Stripe API. Stripe just stayed top-of-stack in a moment of real decision-making.

The agent makes the call, and the product that’s ready to be called wins the moment.

What you should ship and when

The move is simple. Pick one of the three shapes and ship it this quarter. First-class, alongside your web app.

Here’s how to choose between them:

  • If your product already has a public API, ship an MCP server. One to two weeks of focused work.
  • If your product is developer-facing, ship a CLI. Your users already live in the terminal.
  • If your product has no API at all, ship a clean public API with three core endpoints, then put MCP on top.
  • If your product is consumer-facing, this can wait. Revisit when AI agents enter your category in earnest.

Most B2B SaaS founders should start with MCP. The protocol is open, the SDKs are in TypeScript and Python, and a basic server covering your top three workflows is a focused week of engineering time.

It takes less than you think

Here’s the friendly news. Shipping this is small. An MCP server on top of an existing REST API is 30 to 80 hours of focused work. A CLI on top of an existing API is 20 to 40 hours. A clean public API from scratch is bigger, usually 4 to 8 weeks depending on scope.

For most early-stage teams, this is a one-engineer, one-sprint bet. And it opens a channel that compounds. Every new agent that lands (and they’re shipping monthly) finds your product without you doing anything new.

How to make this actually work

The teams winning at this treat the agent layer like a real product surface, not a launch banner. They give it the same care as their web app: clear tool names, sensible defaults, real error messages, rate limits that make sense.

Documentation is the biggest lever, and most teams underbuild it. Agents read your docs as part of almost every call. Sparse docs mean the agent guesses, fails, and your user feels the friction. Rich docs mean the agent picks the right tool on the first try and runs it correctly. Treat your docs as instructions for an AI, not as marketing copy for a human. Be literal. Be specific. Spell out what each parameter does.

Then give the agent examples. Real, working ones. For every tool or endpoint, show three or four common requests with the exact inputs and outputs. Agents pattern-match on examples faster than they reason from prose. A well-chosen example turns a coin flip into a sure thing.

Add a small library of success stories or recipes. “How to do X in three calls.” “How to handle Y.” “What to do when Z.” These are basically prompts the agent learns from. The Stripe and Linear docs are the gold standard here. Read theirs, then write yours in the same shape.

The frame for all of it is simple: every minute you save your user, you keep them. Every dollar you save them, you keep them. Every retry, every failed call, every “I’m not sure how to do that” is a tax you put on the relationship. Good docs, good examples, good recipes remove that tax.

Do that, and something quiet and lovely happens. The agent calls your tool. It works on the first try. The user keeps going. They never leave Claude or Cursor, and they never had to. You just earned a habit without ever showing up.

The shift in one line

The unit of access used to be a click. Then a tap. Now it’s a sentence to an agent. The founders who ship for both surfaces get pulled into more workflows, more recommendations, more daily use, with no extra marketing spend.

The interface layer is plural now. There’s a new channel waiting for you, and it’s wide open.