A directory has two levers: how many entries it has, and how much it can vouch for each one. Pulling the first is easy. Pulling both at once is the entire problem.
Our split
**470 MCP entries.** 127 reviewed — a person read the documentation, wrote the description, and dated the check. 343 listed — compiled from npm, evidence is mechanical: the package exists, declares an executable, and names MCP in its own description.
That is 27%. It was 100% for the first ninety entries, and it will keep falling unless we spend every session on review instead of coverage.
Why not review everything
Reviewing 343 more entries at the current pace is months of work, and the long tail does not deserve it. Half of those packages are used under a thousand times a week. A person's hour is better spent on the server people will actually install.
So the rule is: **review what gets chosen, list what exists.** The reviewed entries are the ones in comparisons, in best-of lists, and in the alternatives pages. The listed ones live in one table and never appear as recommendations.
What the two tiers can and cannot do
A listed entry can tell you a package exists, installs, and roughly what it does. It cannot tell you whether the description is accurate, whether the maintainer responds, or what happens when an agent guesses wrong.
A reviewed entry can tell you all three, because someone checked. It also carries a `bestFor` and a `notFor`, which are the two sentences that require actually understanding the tool.
The label is the product
We could stop labelling and let people assume everything is checked. That would raise conversion on the table page and destroy the only reason to trust the reviewed tier.
The label is not a disclaimer. It is the feature: **the one thing a reader can rely on without verifying it themselves is which claims have been verified.**
What would change our mind
If the data shows the listed tier is what brings people in and the reviewed tier is what they use, the ratio is right. If the reverse, we should cut the table rather than grow it. That is a question for analytics, not for conviction — and it is the first thing we will look at after launch.