How a Simple MindMap App Landed Us in Anthropic’s Trending Connectors 

Platforms like Claude and ChatGPT occasionally open up new ways for businesses to reach potential customers. At Keenethics, we like to understand how these opportunities work while they’re still new, and what value they can bring our customers. With that mindset, we kicked off an internal experiment with Claude connectors.

We were on a call about something else, looking at analytics for a completely different site, when the numbers stopped making sense. There were 5.19 million requests in fourteen days to a domain nobody was promoting or running ads for.

“We were checking on something unrelated, and I said, wait, why do we suddenly have so much traffic here? At first I thought it was a mistake in the report.” — Ivan Goncharov, Head of R&D at KeenEthics

It wasn’t a mistake. Digging in, the team traced the spike to a small app, a mindmap generator we’d built weeks earlier and submitted to Claude as an experiment. It had quietly gone live, sat untouched for a while, and was now pulling in requests.

Somewhere in the noise of app submissions, something had clicked, and nobody had been watching for it. That accidental discovery is the whole reason this article exists.

Why we built a mindmap app in the first place

When ChatGPT launched apps (now plugins), and Claude introduced connectors, both built on MCP, there was a short period when everyone was still figuring out what worked. We wanted to be inside that window with something real.

So we asked a simpler question than “What’s a good app idea?” We asked what kind of app could work for almost anyone, regardless of their job or industry, the same way a flashlight app made sense on every phone even before anyone needed one. 

That meant looking for something with:

• A broad, general audience.
• One clear, obvious use case.
• A simple name and a clear purpose.

Mindmaps checked every box. People build them for notes, planning, brainstorming, studying, project structuring, basically any moment where thoughts need to turn into a shape. And critically, almost nobody had built a connector called simply, “MindMap.”

We didn’t try to make it impressive. One developer put together the MVP in about two and a half weeks, enough to generate a clean, usable mindmap from a prompt.

“We weren’t trying to build something fancy with complicated functionality. We were trying to get onto a new platform fast with a good enough product that could go viral and help us learn how the platform works.” — Max Savonin, CEO and Co-Founder at KeenEthics

We submitted it to Claude and to ChatGPT and moved on to other work. The review took roughly two months before it went live. By the time Claude published it, we’d mentally filed it under “shipped,” closed the loop, and stopped watching it. Which is exactly what made the traffic spike so disorienting when it showed up.

Claude quietly turned on a recommendation engine

We knew that, unlike ChatGPT*, Claude had a recommendation feature that could suggest relevant connectors directly to users, even if they’d never installed them before. What we didn’t know was how much traffic that could actually drive.

Once we started paying attention, the pattern made sense. It wasn’t coming from people browsing a marketplace or searching for connectors by name. ChatGPT plugins sit in a more discoverable spot than Claude’s connectors do, and ChatGPT has a much larger user base. Yet the traffic there was much smaller, roughly 200 new users a day, nothing like what we were seeing on Claude.

Here’s how the feature, still in beta, works in practice: 

1. Someone asks Claude for something, say, “build me a mindmap of this project.”
2. If they don’t have a relevant connector installed, Claude can surface from one up to four** options on its own.
3. The person picks one, connects it in a click, and the connector creates a mindmap.

It was purely a recommendation experience. We previously called it indirect prompting, but we’ve since separated the two terms. Indirect prompting is when Claude offers to use an installed connector, while Recommendation is when Claude suggests one to four new connectors to use.

“Claude was recommending our connector to people who hadn’t installed it yet. In other words, it was introducing new users to the connector, not simply using a tool they already had.” — Ivan Goncharov, Head of R&D at KeenEthics

Claude can recommend installing a new connector or use one you already have, depending on what helps answer your request. ChatGPT experimented with something similar that terminated because users read it as advertising and now in September 2026 they turned it on back again***.

Which means the millions of requests we were seeing weren’t coming from returning users. We didn’t have an audience yet. This was all new traffic, generated by Claude recommending us to people who had no idea we existed.

The numbers, roughly

We already knew the 5.19 million figure from that first call, but we sat down and pulled up the actual dashboards to see the fuller picture.

Here’s what the analytics showed:

~5.12 million requests hit the app’s logo endpoint.

~414,000 requests hit the endpoint where Claude asks “what can this tool do,” meaning that many times someone (or Claude on their behalf) was curious enough to check.

~861 mindmaps were actually generated.

That first number is almost entirely one thing. We’re about 95% confident it reflects Claude pulling our icon every time it recommends our connector, back when the list still topped out at four options. In other words, most of that traffic wasn’t usage at all. It was Claude showing people our name and logo, over and over, as a suggestion.

The 861 number is the one that stung a little. Even accounting for the gap between “shown a recommendation” and “installed the connector,” we’d expect far more than 861 mindmaps from those installs. They didn’t materialize, and the reason turned out to be on us. Once someone installed our connector, Claude just didn’t call it to draw the map.

We wanted to publish as fast as possible, mainly to see how Claude would interact with our app and start refining our tool descriptions from there (something we paid much more attention to with Real Random). But as early submitters, we went through a Google form, which let us publish the connector but gave us no way to edit it afterward.

Despite staying in touch with Anthropic’s team, we still don’t have dashboard access today. If you publish through Claude now, that access comes right away, so you can adjust your tool descriptions as you go. The tool itself is called render_mindmap, but people usually ask Claude to “create” or “build” a mindmap, and Claude has to match that intent to a tool description that, in our case, was quietly working against us.

On the ranking side, we spent 13 days climbing Anthropic’s trending connectors list, moving from around 25th place up to 3rd before things leveled off.

That same lack of dashboard access meant we also couldn’t cross-check any of this against Anthropic’s official numbers. Everything above comes from our own Cloudflare logs on the code domain.

Why our small app got picked so often

Once we started digging into how often we showed up, the story got more surprising.

When someone asks Claude for a mindmap and doesn’t have a connector installed, Claude picks a small handful of options to recommend. In this category, Figma and Miro sit at the top of it, and there are roughly 15 smaller competitors beneath them. 

We ran our own informal test, with dozens of prompts and attempts, phrased the way a real user might ask for a mindmap. Out of all of it, we saw Claude recommend our app exactly once.

That’s part of why the traffic still catches us off guard. Recommendations are in beta, they don’t fire on every relevant request, and they’re rolled out to a limited slice of users. But unlike ChatGPT, Claude checks its connector directory for a relevant match before answering certain requests itself, rather than just answering from its own knowledge.

We rarely landed in the recommendation list at all. And yet, somehow, that one narrow window was still enough to generate the numbers we saw, which says more about the size of Claude’s user base than it does about how often we were actually chosen.

“We didn’t win a competition. Most of the time we weren’t even in the running. But apparently you only need to get picked a handful of times out of millions of conversations for the numbers to look enormous.” — Max Savonin, CEO and Co-Founder at KeenEthics

Being in a smaller category and getting listed early both help your odds, that part’s real. But getting there early only takes you so far, because what happens next depends on the groundwork. What decides whether Claude reaches for your connector is the quality of your tool descriptions and metadata, the exact wording that tells Claude what your app does and when it’s the right fit. 

Fewer competitors buy you a chance, but well-written metadata is what turns that chance into something Claude reliably picks.

Where we go from here

In this story, we’re sitting with 5.19M requests, and only 861 mindmaps to show for it. We don’t fully understand that gap yet, and we’re not going to pretend we do. But we know enough about the mechanics to have a real theory.

Claude decides which connectors to recommend based on three things: 

1. The connector’s overall description.
2. The tool’s name.
3. The tool’s own description. 

That gives us a pretty good idea of where things may have gone wrong, but the bigger takeaway goes beyond one small app. We spent a year experimenting with MCP, building things, watching the ecosystem take shape, largely in the dark about whether any of it would translate into real reach. 

This is the first time it did. It didn’t come from SEO, which takes years to pay off, or from an app store listing. It came from a platform with millions of users doing the discovery for us, and putting our app in front of people who were already asking for exactly what we built. That’s a different kind of channel than anything we’ve used before.

That won’t stay this easy forever. Right now, this feels a little like the early web, when a mediocre site with the right keywords could rank near the top. But we don’t think that phase will last. AI systems are already getting better at telling what’s genuinely useful apart from what’s simply optimized to look that way, and we expect that gap to widen.

As more low-effort connectors get built to chase this mechanic, clear descriptions and real value after someone connects will matter more. 

We suspect that if our own tool-calling had worked cleanly, taking people from a recommendation to an actual mindmap, we might be sitting in that trending list today.

“We stumbled into this one. Next time, we want to walk into it on purpose.” — Ivan Goncharov, Head of R&D at KeenEthics

*ChatGPT turned the app-suggestion feature back on for all users, not just early partners, in September 2026.
** In summer 2026, it was up to 4, and in September 2026 it can be up to 5.
*** According to TechCrunch, OpenAI paused app suggestions in December 2025 after ChatGPT Plus subscribers called them ad-like, some had seen unrelated conversations interrupted with promotional mentions. It left indirect prompting with installed plugins in place and has since started recommending individual plugins again in more limited cases.