AgentCat: the silent failures of MCP servers

Kashish Hora works at AgentCat. Luca Berton talked to him at AGNTCon + MCPCon North America 2026 in San Jose on 23 October 2026, at an AgentCat booth covered in cats. A few minutes into the first take he found he was allergic to them, and they started again from the top.

Easy to build, hard to understand

AgentCat, as he puts it, helps companies building MCP servers “understand how agents are using those MCP servers, where they’re getting stuck, and how they can fix it.”

Building an MCP server is really easy nowadays, he says: there are many tools to get started, and the MCP developer community has made very good SDKs. “But once you’ve released your MCP server, something that’s still really challenging is understanding how agents are using it beyond just logs and traces.”

Silent failures

What AgentCat looks for is what it calls silent failures: an agent uses your MCP server to accomplish something and can’t. Maybe it gives up, maybe it calls the wrong tool and gets confused. Observability and telemetry tools capture logs and traces, and those catch the typical errors, an authentication error for instance. “You won’t see when an agent gives up or abandons the session early in your logs and traces.”

Take an abandoned session. An agent tries to do something for its user, but the MCP server lacks the functionality or it is hard to find, so the agent gives up, and the user concludes the tool cannot do it.

One line on the server

AgentCat has SDKs in all the major languages for MCP, TypeScript, Go, Python and Ruby among them, and installing one takes a line of code that wraps the MCP server. Behind the scenes the SDK works almost like middleware: it improves the tools and adds information the agent can provide, and the platform aggregates it all for the developer. Among other things, it lets the agent report back what it could not do, so the developer collects that feedback over time and improves the server.

Luca asked whether that is auto-instrumentation. Exactly, he says: the SDK instruments the server with everything the product needs, and the data it collects can optionally be sent on to any OpenTelemetry collector or receiver. “Our goal is not to compete with your existing analytics and observability stack.” Nor is it an API gateway, a question AgentCat gets a lot: “Our product lives entirely on your MCP server.”

The agent experience

MCP comes first because it is the most common and widely adopted way companies build agent experiences. For two decades, companies built websites and web apps around people looking, clicking and typing. “We believe that the next two decades, the most successful companies will be the ones that prioritize how agents experience your product.” Agents already try workarounds, opening websites and clicking buttons meant for humans, and they “will be the biggest users of software products in the next 20 years.” AgentCat has work under way for WebMCP, for basic browser use, and for standards such as A2A and UCP.

Luca recalled a point from Nordic APIs: APIs embed a workflow, log in first and then act, with more for premium users, while an agent tries every possible way, with no guided path. “It’s a different way of thinking about software development,” Kashish says, a new muscle for many organisations, with new roles such as head of AI and head of agent experience.

Webflow, and how to start

One of AgentCat’s biggest customers is Webflow, the AI-native website building and content management platform. AgentCat has helped it as its MCP server grew, and in making its whole product suite better for agents.

The product is fully self-serve, free to try and free to use, and all the SDKs are open source. The place to start is agentcat.com.

More

The episode with Kashish Hora is being produced and goes out on YouTube first. The day it was recorded: AGNTCon + MCPCon North America 2026: day two.