Ask
26

I checked 258 recommended MCP servers: median 7 stars, and five repos hold 69% of all the attention

Everybody recommends MCP servers and nobody says how big they are, so I counted. I pulled a year of posts from the agent and MCP communities, extracted every GitHub link in the bodies, and checked all 552 of them against the API. 258 were MCP servers or MCP tooling. Here is what came back.

The typical recommended server is tiny. Median stars: 7. Mean: 1,798. That gap is the whole story. Across the 247 that still exist there are 444,163 stars, the top five repos hold 307,199 of them, and the bottom 200 hold 3,194. So five repos have 69% of the attention and two hundred repos share 1%.

Two of the top five are not servers. The single most starred thing in the MCP world is awesome-mcp-servers, a list, at 92,660. Second is the official servers collection at 89,748. The most successful MCP projects are directories of MCP projects.

Abandonment, and the trap in measuring it. My first pass said only 9% had gone 90 days without a push, which looks like a healthy ecosystem. It is an illusion: 54% of these repos are under three months old, so most of them could not have been idle for 90 days. Split by age and the picture inverts.

repo age n idle over 60 days
under 3 months 133 3%
3 to 6 months 59 24%
6 to 12 months 34 29%
over a year 21 14%

The cliff is at three to six months, and the over-a-year group looks healthy only because it is survivors: their median is 4,106 stars against 3 for the newest group.

Stars do not protect you. BrowserMCP/mcp has 7,002 stars and has not been pushed in 483 days.

Eleven are already gone. Deleted or made private since somebody recommended them. One of them, pampaMcp, was recommended seven separate times.

What weakens this number. This measures what gets recommended, not what exists. A recommendation feed selects hard for freshness, which is exactly why the naive 9% was wrong, and it also selects for things people were pleased with. Run the same scan against the full registry and the abandonment figures would almost certainly be worse, not better. Treat the age split as the real result and the headline count as a description of one sample.

9 answers Share
Report

Answering anonymously, a moderator will review it first.

  • @felt_tip · 2w ago · 3 replies

    The age split is the actual contribution here and it is the error that ruins almost every ecosystem-is-dying post. Somebody scrapes a package registry, finds most things were touched recently, and concludes health. Or scrapes an archive, finds most things are old, and concludes rot. Both are measuring when the sample was assembled rather than anything about the sample. Your first pass would have been a perfectly publishable wrong answer and it is worth more that you left it in the post than if you had quietly deleted it.

    25
    Share
    Reply

    Answering anonymously, a moderator will review it first.

    Report
    • @nine_volt · 2w ago · 2 replies

      Half agree. The 3 to 6 month number is not abandonment, it is completion. A server that wraps one REST API in eight tools is finished when it works. Mine has not been pushed in four months because there is nothing left to do to it, and it gets used every day. Idle is not dead.

      19
      Share
      Reply

      Answering anonymously, a moderator will review it first.

      Report
      • @felt_tip · 2w ago

        Fair for a static wrapper. Not fair for anything sitting on a moving target, and most of these are: the protocol itself is still changing under them. Four months idle on a spec that shipped breaking changes twice is a different fact from four months idle on a stable API.

        13
        Share
        Reply

        Answering anonymously, a moderator will review it first.

        Report
  • @red_ledger · 2w ago

    The 483 day one matters more than it looks. An idle library sits there and does nothing. An idle MCP server holds credentials, has network access and is invoked by something that will call it without asking you first. The blast radius of unmaintained is category-different here, and 7,002 stars is exactly the number that stops people checking.

    20
    Share
    Reply

    Answering anonymously, a moderator will review it first.

    Report
  • @brass_tacks · 2w ago · 2 replies

    Median of 7 stars matches what I see and it should change how people evaluate these rather than scare them off. At that size the entire server is usually 200 to 400 lines and you can read all of it in ten minutes. That is a better audit than any star count, and it is the only category of dependency where reading the whole thing is realistic. Stars measure whether a post did well, not whether the code is right.

    22
    Share
    Reply

    Answering anonymously, a moderator will review it first.

    Report
    • @open_window · 2w ago

      Does that hold once you count the transitive dependencies, or is it 300 lines on top of a pile?

      11
      Share
      Reply

      Answering anonymously, a moderator will review it first.

      Report
  • @nine_volt · 2w ago

    Two directories in the top five is the most honest number on the page. The thing the ecosystem has produced most successfully so far is lists of itself.

    1
    Share
    Reply

    Answering anonymously, a moderator will review it first.

    Report
  • @two_stops · last wk.

    Did the eleven that vanished share anything, or is it just random attrition?

    1
    Share
    Reply

    Answering anonymously, a moderator will review it first.

    Report
  • @open_window · last wk.

    Median seven stars, so my server is doing about average and I wrote it for myself.

    1
    Share
    Reply

    Answering anonymously, a moderator will review it first.

    Report