Case study · Webflow
How Webflow builds its AI roadmap based on actual agent behavior
Webflow's MCP server took off the week it launched. Here's how they used AgentCat to turn raw agent traffic into the data behind their AI roadmap.
110%
month-over-month growth
Days → hours
to debug an issue
54%
of sessions fall into 3 use cases
“AgentCat changed how we make decisions. We share screenshots of graphs on the Dashboard internally, pull up Agent Goals in our weekly planning meetings, and walk through Session Replays every time there's a bug.”
Impact
- Classified agent goals and intent across all sessions at enterprise scale
- Reduced debugging time from days to hours
- Detect issues before customers report a bug
- Identified a single silent error affecting 15% of all sessions
- Discovered 54% of sessions fell into 3 use cases
- Saved countless hours a week by automating all data collection and classification
- Highlighted top use cases to help prioritize AI roadmap
About Webflow: Webflow is the agentic web marketing platform that empowers modern marketing teams to build, manage, and optimize websites and web apps with the speed, scale, and intelligence that today's brands demand. More than 300,000 global companies like DocuSign, Upwork, Dropbox, and Orangetheory Fitness use Webflow to deliver fast, flexible, and high-performing brand experiences at scale.
When we first met the team, they had just launched their MCP server but were already seeing an exponential increase in usage. We were excited to help Utkarsh Sengar, Webflow's VP of Eng, understand how users were using their MCP server, see where it was falling short, and utilize those insights to prioritize their AI roadmap for the rest of the year.
THE PROBLEM
A fast-moving team building for agents, and existing tools just weren't cutting it
Webflow built the first iteration of its MCP server in just a week, and immediately started seeing interest. Users were requesting features as fast as Webflow could ship them. Their MCP server let agents build responsive page layouts, manage content collections, apply and refactor styles, optimize for SEO, and run quality checks all through natural language.
Their existing tooling provided logs and traces, but that only showed when individual tool calls errored. It failed to help them understand what agents were doing and where they were getting stuck.
At the same time, MCP, as a protocol, was evolving rapidly, and AI agents were becoming more advanced. With usage of their MCP server increasing every day and the user requests piling on, they didn't have time to spare. Webflow had to prioritize fast.
"We were prioritizing our roadmap based on traces and logs, not real insights. That's all we had with our current tools."
From a few conversations with their early users, the most talked about use case seemed to be managing CMS, such as blog posts and storefront listings and migrating from other platforms like WordPress.
The team was planning a huge Q1 push to expand CMS tooling in their MCP server. It matched what they'd heard, but they had no way to validate that assumption with data.
THE SOLUTION
The first two weeks of data changed the roadmap
AgentCat's SDK took just one line of code to install, and Webflow automatically started seeing end-to-end sessions: which tools agents called, why they called those tools, whether any errors occurred, and what agents were ultimately trying to accomplish. For the first time, the team could see intent, not just activity.
But the turning point was Agent Goals.
The Webflow team already knew that a major use case for the MCP server was large-scale data migrations from legacy platforms like WordPress and Adobe. But AgentCat's Agent Goals revealed a new dominant use case: using third-party agents to design in the Webflow canvas, accounting for over 24% of all sessions.
The team quickly prioritized improving the MCP server to handle this rapidly growing use case. One week of insights through AgentCat helped them prioritize their entire quarter.
And they didn't stop there.
They started incorporating AgentCat into their product development workflows. The Dashboard page gave them a high-level overview of how their server was growing. When a user reported an error, the team could go right into the Session Replay to see exactly what happened. And they could see which errors were affecting the most users on the Issues page to prioritize fixes by impact.
Buried in the session data was something no one expected: agents were calling tools that didn't exist. Agents were guessing at names, trying variations, attempting to discover capabilities through trial and error. This wasn't a bug. It was thousands of agents saying, "I can't find what I'm looking for." Traditional monitoring would log these as errors and move on. AgentCat surfaced them as a pattern and turned it into a documentation and discoverability roadmap that the team didn't know they needed.
RESULTS
With the help of AgentCat, the Webflow team was able to identify use cases, prioritize their roadmap, and quickly debug issues.
Webflow's MCP server has grown 110% month over month since launch. AgentCat has helped them make sense of that growth by identifying agent goals for all sessions and reducing debugging time from days to hours.
Going forward, Webflow is using AgentCat to make its existing data more valuable.
They're planning for a future in which AI agents are one of the primary users of their products. They're optimizing their product surfaces to be agent-first, and AgentCat is at the center of that change.
"We already have a wealth of data that powers Webflow in a data warehouse, and are constantly figuring out how to use that data to build better products and features for our users," says Utkarsh.
"AgentCat's platform has already become a crucial part of our workflow. But as we continue building agent-first experiences, being able to tie AgentCat's intelligent agent analytics into our backend data is going to be a huge unlock."