Case Study

When your users are AI agents,
they don't complain. They just churn.

Exa's MCP server was handling millions of agent calls with no visibility into what was actually happening. Agnost changed that.

14h to 3m To debug an MCP incident
Instant alerts On error spikes, tool failures, and bad call patterns
About

Exa is a search API built for AI agents. Developers plug Exa's MCP server directly into their agents to give them real-time web search, crawling, and research capabilities. When it works, it's invisible. When it doesn't, nobody says anything.

The Problem

Exa's MCP server was growing fast. Developers were building with it, agents were calling it millions of times, and the product was clearly resonating. But Exa had a blind spot that every MCP-native product eventually runs into.

Their users were AI agents. AI agents don't leave reviews, file support tickets, or tell you what they need. They just fail silently and move on.
When something went wrong, they found out one of two ways: a developer complaint, or a manual dig through noisy logs.
No structured view of which tools were underperforming, which query patterns were causing failures, or which developers were quietly having a bad experience.

They were building a product used by millions of agent interactions with almost no insight into any of them.

The Solution

Exa connected Agnost to their MCP server. For the first time, every tool call had context, and every pattern had a name.

Tool-level visibility Which tools were underperforming, on which query types, with what error distribution. Not just "something's broken" but exactly what and why.
Developer interaction analytics A structured view across every session: what developers were actually trying to do, where they were struggling, and where the MCP was letting them down.
Proactive alerts Daily summaries, Slack alerts, real-time signals. Instead of waiting for a complaint, Exa gets flagged the moment error rates spike or a bad experience pattern emerges.
Failures, caught in session When something breaks, they see it in the session it happens. Not in a support email days later.
The Result

A 14-hour debugging session became a 3-minute one. That's the incident that made it real.

When something breaks, they see it in the session it happens. When developers are struggling with a particular tool or query type, that pattern surfaces before it becomes churn.

Every product decision is now grounded in actual usage. Exa knows how developers are using their MCP server, what they're building with it, and where things need to improve.

"We used to find out something was wrong when a developer told us. Now we know before they do. Agnost gave us the product analytics layer we didn't know was missing."
Ishan Goswami
Ishan Goswami Member of Technical Staff, Exa
Ready to see how your agents are performing?
Get Started or Book a Call