What an agent buys is context, not tools

Prowl is one MCP endpoint that hands an agent the research it would otherwise assemble by hand: SEO, ads, SERP data, reviews, trends. Every surface I own has printed the same figure about it for months, 408 tools and 12 modules^E2, and today I asked the endpoint instead of the page. It answers 448 tools in 8 categories^E1. The catalogue grew, the copy didn’t, and that is the cheap half of the mistake.

The expensive half is that my own registry already carried this row marked needs primary source, with the check written beside it in one line: take the count from the live endpoint, not from the marketing page^E5. That line had been sitting there twelve days (the staleness rule on that file allows ninety, which is the second thing this catches). What does it cost to run? One call, and nothing.

What an agent buys here isn’t the 448 anyway. It’s the eight playbooks, and each one names who it is for in its own metadata^E6: idea-validation for a founder holding an idea with no domain yet (competitors, sizing, channels, a verdict, six sections), channel-economics-audit for whoever has to find where attention is under-priced, competitor-teardown for an agency that needs to walk in with the teardown already done. A catalogue is inventory. A playbook is the thing somebody was otherwise going to assemble by hand on a Sunday.

So what does an agent pay to walk in the door? The price of an agent-facing product is context, which is easy to miss if you have only ever priced things in dollars. The first call an orienting agent makes is some version of what is in here, and that call used to return all 448 names: 16,072 characters, about 4,018 tokens^E3. It now returns eight counts and a hint, 361 characters^E4, and hands over names when somebody asks for them. Same catalogue, 2.2% of the entry price, and you can run both calls yourself in about two minutes.

But none of this tells me whether anyone runs those playbooks. I have a catalogue count and an entry price, both read from the endpoint an hour ago, and no usage curve I would put in front of you (the endpoint answers what it holds, never who called it). A post about a product with no retention number in it is a post about the build, and this one is.

The correction at least runs in the pleasant direction: the figure was wrong low, by forty tools. It gets fixed from the endpoint rather than from the page, in five places including my own table of good examples, where 408 market-intelligence tools behind one MCP endpoint had been standing as the model of a claim with a source behind it. That row has one now.