Half of the MCP packages on npm are used fewer than a thousand times a week
We measured 343 packages. The median gets 634 downloads a week, 55% are under 1,000, and 41 recorded none at all. What that distribution means for anyone choosing a server.
Field notes
What we learn while checking listings: which tools hold up, how to spot an abandoned project, and what this directory actually earns.
We dropped 44 packages that were not MCP servers
Searching npm for 'claude mcp' returns Playwright, Clerk, and Cline. Here is the evidence rule we added, and the 11% of our own data it removed.
56% of the MCP servers we reviewed are run by the vendor
71 of 127 reviewed entries come from the company whose product they expose. Official status changes what you should expect — and what you should not.
71% of MCP packages are MIT — and that tells you almost nothing
245 of 343 listed packages carry MIT, 51 Apache-2.0. What the license does and does not say about whether you can rely on the thing.
42 of 46 AI tools have a free tier, and that is a pricing signal
Almost every tool in the directory can be evaluated without a card. The four exceptions tell you more about the market than the 42 do.
Why only 27% of our listings are reviewed
470 entries, 127 checked by hand. The ratio is deliberate, and it is the number we are least willing to inflate.
What the verification date on a listing actually means
Every reviewed entry carries a date. Here is exactly what was checked, what was not, and what happens when a date gets old.
What we refuse to publish
No traffic estimates, no fabricated star rankings, no paid placement inside rankings. The exclusions define the site more than the inclusions.
Skills are not plugins, and the difference decides what you install
35 skills in the directory, and most of them change how an agent behaves rather than what it can do. That distinction is the one users get wrong.
How we decide which tools get compared
21 head-to-heads out of thousands of possible pairs. The filter is not popularity — it is whether a real decision exists.
The tutorials exist to explain the directory
18 lessons, four tracks, no growth-hacking content. Each lesson ends where a tool page begins.
What 156 MCP packages on npm actually look like
We compiled every npm package that declares an MCP server, then measured the distribution. The median package gets 4,517 downloads a week. Thirty-four get none.
An npm keyword is not a directory
Searching npm for 'mcp' returns libraries, adapters, proxies, and a few things that are not servers at all. Here is how we tell them apart, and what we refuse to list.
Debugging an MCP server locally in five minutes
Most MCP failures are not model failures. They are a server that never started, a transport the client does not speak, or a tool that returns an error string the agent reads as an answer.
Your agent is reading the wrong file
When an agent edits the wrong code, the reflex is to blame the model. Usually the context was wrong before the model ever saw it — and that is a tooling problem, not an intelligence problem.
What your tool definitions cost before the conversation starts
Every MCP server you connect adds tokens to every single request, whether or not the model uses it. Ten servers can eat a fifth of a small context window before the user says anything.
How to tell an MCP server is abandoned before you install it
The protocol shipped fast, the directories filled up faster, and a lot of what you find in a search result has not seen a commit since last winter. Five checks, in order of how fast they rule a project out.
The AI coding tools that survive the first 30 days
Most of the tools in this category are impressive in a demo and gone from your workflow by week three. The difference is not model quality — it is where the tool sits in the loop you already have.
A directory without a verification date is a rumour with better typography
Every AI directory claims to be curated. Almost none of them can tell you when the price it published was last true.
What an AI directory actually earns from AdSense, with the arithmetic
Publishers talk about RPM like it is a single number. It is a range that depends on geography, placement, and what the page is about — and the range for developer content is wider than most people expect.