Affogata Benchmarks - Live Data Explorer

Dating app report made by AI, entirely off the data via WebMCP
Deep dive into a finding from the data about Bumble - again purely via WebMCP
Agent independently picks the fastest-climbing app + its two closest rivals, writes them into the live comparison, and gives a real verdict
Compare view empty, before the tool call
Same view after `benchmarks_compare_in_ui` fired (with Bumble/Hinge live on screen and the tool-call log visible)
The tool trail next to the rendered brief, shows the full call chain in one frame
Working with the app from the ChatGPT app, just the data
Working with the app from the ChatGPT app with the app visible

Inspiration

Affogata already analyzes & scores dozens apps and games across multiple categories, analyzing real conversational data from social and review sites that are often hard for LLMs to access. This data lives behind a login and your usual "SaaS app" UI, where business users rely on it (from our customers such as Mojang, EA, Scopely and others), or separately an agent could connect to it via MCP/API.

But how do you bridge these worlds? AI working with data is magical, but people do need the visuals and exploration to get a sense for the data! When we saw what WebMCP can do, it looked like the natural evolution for human + agent collaboration. Instead of a siloed experience where humans use a UI and agents use a backend connection to look at the same data in different ways, a singular experience can offered where one truly complements the other and natural exploration and "experience" of the data can happen.

What it does

benchmarks.affogata.net exposes 15+ WebMCP tools, both reads and a write, directly on the live site with a 79-title portfolio of real brands. The tools include:

benchmarks_get_overview, benchmarks_get_movers, benchmarks_get_title, benchmarks_get_category_analytics, benchmarks_list_titles, benchmarks_compare_titles, benchmarks_explain_methodology - reads
benchmarks_compare_in_ui, benchmarks_open_title - writes that change what's rendered on the page for anyone watching, not just inside the agent's own chat

An agent using nothing but these tools can diagnose why a title's score dropped, compare it against its closest rivals, put that comparison live on screen, scan the entire portfolio for the next brewing crisis, analyze the data and prepare reports, all without a single click or scrape, and with a experience where the AI working with the data actually improves what the human is seeing on screen.

How we built it

The benchmark scoring already existed as Affogata's own MCP schemas. We adapted the ones that map to public, read-only data into the WebMCP registration format and added one write tool (benchmarks_compare_in_ui) so an agent can drive the compare view live, not just describe it in chat. The site runs the tools in page JS against the browser session, so there's no separate auth or API key: An agent using ChatGPT's in-app browser or Chrome with the WebMCP flag on gets the same tool list a logged-in human would see.

Challenges we ran into

Getting an agent to reliably discover and call the right tool names instead of guessing from page copy took a couple of passes; an early bug had benchmarks_compare_titles and benchmarks_compare_in_ui named inconsistently between the page text and the actual registration, which caused the agent to call a tool that didn't exist. We also hit a live-state sync issue: when a human removed a title from the compare view by hand, the agent's next tool call still answered from memory instead of re-reading the page; exactly the kind of human-and-agent-together moment WebMCP is supposed to get right, so it was worth fixing rather than working around. Finally, the schema needs to be clear to models lower on the tool-using skill rank that the user might be using if they're using Chrome, so there was some iteration on that as well.

Accomplishments that we're proud of

This app has real utility! It's a vision for data apps that work great for humans and agents at the very same time. To that end, we've built three end-to-end demos built entirely from tool calls, no manual data entry: a single-brand "Customer Goodwill Crisis Brief" that diagnosed a real -4.8 score collapse and generated a branded, print-ready HTML report with every number traced back to a named tool call; a portfolio-wide "War Room" that scanned all 79 titles at once, ranked them on a transparent risk index, flagged a false alarm, and surfaced an improving challenger; and an agent that independently picked out the fastest-climbing app, pulled in its two closest rivals on its own judgment, put that comparison live on screen, and gave a real analytical verdict on whether the gain was sustainable; the kind of judgment call that used to need a person opening a dozen tabs.

What we learned

The write side of WebMCP is what actually changes the human-agent relationship. A read-only agent is still just answering questions in a chat window. The moment benchmarks_compare_in_ui puts a comparison on screen live, two people can watch the same tab while an agent drives it; that's a genuinely different way of working together than screen-sharing a chat. It's great for a person working with the data; it's even better when a team wants to have a live discussion informed by it.

What's next for Affogata Live Data Explorer

Wire the demo tool set up to Affogata's live customer data (today it runs on weekly-refreshed benchmark data), and bring WebMCP support into the logged-in Affogata product itself, so any customer can point their own agent at their own data without an admin setting up API access first.