What shipping an MCP server taught us about our customers 

When Anthropic announced MCP, we weren't sure it would become the standard. But it was clear that how people would interact with software like ours had fundamentally changed. The era of "log in, click around, build" wasn't ending, but it was no longer the only way in. Agent integrations were coming. Our customers would expect them. We needed to be ready. What we didn't anticipate was who would show up first.

Our MCP journey started as a 1-person, 5-day sprint in April 2025, a bet on a protocol that only one major vendor was backing at the time, and a calculated decision to move before we knew if it would stick.

It turns out, it stuck.

Customer.io is for marketers, generally startup to mid-market, data-driven teams who want to send the right message at the right time. That's our core.

But within weeks of shipping our MCP server, a new kind of user had found Customer.io: developers, solo founders, tinkerers building one-person operations with AI tools stitched together. They weren't the marketers we'd built for. They were reaching into our platform through Claude and Cursor, running workflows we hadn't designed, solving problems we hadn't imagined. MCP opened a door, and people we hadn't planned for walked right through it. That surprised us. And it changed how we thought about almost everything.

At first glance, that might sound like noise, like the low end of the market. We didn't see it that way. These are the people operating at the edge of what's possible right now. And a one-person operation running on Customer.io and Claude today could be a 50-person company in two years. More importantly, they're telling you something about where your whole market is heading. We pay attention to them.

What they taught us was that the bar for what a small team can accomplish has fundamentally shifted. AI isn't just automating tasks; it's compressing the gap between a scrappy startup and an established marketing operation. We believe that's an expansion of our market; we welcome those forward-thinkers.

Almost instantly, their AI agents handed us a list of 20 capability gaps we had never prioritized. We discovered them not because customers told us, but because we watched those agents try to use our product and fail.

The first thing customers wanted wasn't to do more. It was to understand how they were doing. Are my campaigns working? Am I actually moving business metrics that matter? Our platform can be complex, and customers had set it up in complicated ways that fit their needs. They were looking for AI to help, so we built performance analytics reporting so agents could surface engagement data. And we launched our Goals product to tie campaign performance to actual business outcomes, sign-ups, conversions, and retention.

Customers quickly moved from “tell me how I am doing” to “do it for me”. That's the shift from AI as assistant to AI as actor. It reshaped how we think about our product roadmap at a fundamental level. Before MCP, our roadmap was organized around UI and product surfaces, around what a marketer could click on, configure, and build. After MCP, every capability had to pass a new test: can an agent reliably, autonomously, and safely invoke this? Those ended up being very different lists.

Here’s something that may surprise you: MCP didn’t replace anything. We didn't add MCP and subtract the UI. We added MCP, then a CLI, then an in-product agent experience. And the UI is still there and essential for the majority of our customers.

The product surface got larger. That's both the opportunity and the hard part.

Different customers are on different parts of an adoption curve. Some may always want a UI. Others want to drive via MCP through Claude. Others—developers—increasingly want a CLI with composable primitives they can script via tools like Claude Code. We want to serve all of them. And they all have to feel like first-class experiences.

This is where I think UX is getting redefined. A well-crafted API or CLI has always mattered to me. I come from a developer tools background. But now it matters in a different way. MCP tools and CLIs need progressive disclosure. They need to be token efficient, surface the right context, and guide agents toward the next best action without overwhelming the model. That's a UX problem, and we're actively working through it.

Our competitive landscape used to be a short list of named SaaS vendors. That list still exists (for now). But it's no longer complete.

The new competition is the internal DIY project, the growth team that decides to build exactly what they need with Claude and some API keys instead of buying a platform. That capability didn't exist two years ago. Today, it's a real alternative for a certain kind of buyer. And it will become more capable over time.

This doesn't mean you can't win. It means your product has to be genuinely easy for agents and tools to integrate with. Because the buyer is now evaluating not just whether they can use your product, but whether their AI stack can use it. A beautiful UI that an agent can't reach is an incomplete product now.

The internal mantra we've been working with: if a person can do it in your product, an agent must be able to do it too. I love the simplicity of it, but it's hard to execute. This is the principle driving most of our engineering and product decisions right now.

We're still figuring this out. Despite every podcaster’s desire to tell the future, nobody can say where this will lead. A 12-month roadmap exists, but it only applies to part of what we're building. We believe the rest has to stay flexible because the current environment changes faster than any planning cycle can capture.

To stay grounded, we keep coming back to a simple question: what do customers need their marketing platform to not simply show them, but actually do? And are we making that possible for both human and agent operators?

That question leads to better primitives, better APIs, better documentation for machines, better guardrails, and better observability into what agents are doing on behalf of customers.

The companies that win this next era won't necessarily be the ones who solve product complexity via fancy UIs. They'll be the ones that become the most reliable execution layer for the marketing workflows that agents are increasingly running. That's the bet we're making.

That 5-day sprint I told you about at the beginning of this post? Eventually, that became what we shipped earlier this month: the AI Agent, Goals, LLM Actions, and more. It's the biggest product release in Customer.io's history. Take a look at what it became.

It started with a 5-day sprint and a protocol nobody was sure would last, and so far, it's paying off.