跳到正文
DEV Community· Andrei·· 55 分鐘前AI 評分60

遠端 MCP 伺服器實測:9/10 可回應,中位數工具定義約佔 4,800 tokens

The real state of live MCP servers: 9 in 10 answer, and what they cost you

AI 導讀

Serverdex 掃描約 21,700 個遠端 MCP 伺服器,發現約 9/10 能完成握手,64% 無需登入、26% 須登入。回答的伺服器中位數提供 14 個工具,定義約佔 4,800 tokens;前 10% 約佔 41,000 tokens,最重的一個約 247,000 tokens。數據來自 2026-11-21 的掃描,僅涵蓋提供遠端端點的伺服器。

正文

Originally published on Serverdex, where these numbers update live.

Every large MCP directory shows you a server count. Almost none of them tell you whether those servers are actually alive.

A listing is a claim. A server that was published a year ago and has been down for months still sits in most catalogues with a green card and an install button. What you actually need to know before you wire it into an agent — does it answer, does it want my keys, how much of my context does it eat — is nowhere on the page.

So we measured it. Serverdex calls every MCP server that exposes a remote endpoint with a credential-free initialize + tools/list, about once a day, and records what came back. Here is the measured reality, from the scan on 21 Nov 2026.

9 in 10 remote servers answer

Of the ~21,700 remote servers we reached:

  • 64% answered the handshake with no sign-in at all.
  • 26% answered after a required sign-in (an OAuth flow or an API key).
  • 10% were unreachable or replied with something that is not the MCP protocol.

So 9 in 10 remote MCP servers that we reach actually answer. The ecosystem is far more alive than the "graveyard of dead servers" framing suggests — but that last tenth is exactly what a directory should flag, and most do not. You can see the servers answering right now on Serverdex's live list, or run the check on any single server with the MCP health check.

(Servers that only ship as a local package — an npm or PyPI install with no hosted endpoint — can't be handshaked from the outside, so these liveness figures are about the ones that run remotely.)

More than 1 in 4 ask for your keys

About 28% of the servers that answer require a sign-in before they expose a single tool. That is not a red flag by itself — it is what an auth or a deploy tool does. But "needs my credentials" is the first thing worth knowing about a server, and almost nobody checks it before connecting.

The real risk with an MCP server was never a hidden instruction buried in a tool description. We scanned ~310,000 tool definitions for that and found it is effectively nil (here's that scan). The risk is capability: a tool with real reach that you wired into your agent without looking. Whether it needs your keys is the first and cheapest thing to vet.

The median server costs you ~4,800 tokens

Every connected server's tool definitions are loaded into the model's context before it reads your prompt. That is a cost on every turn, and it is wildly uneven:

  • The median answering server exposes 14 tools and loads about 4,800 tokens of definitions.
  • The top 10% load about 41,000 tokens.
  • The single heaviest loads about 247,000 tokens — more than a whole 200K context window, from one connection, before you ask it anything.

The full distribution, and how to keep an agent lean, is in The MCP context tax.

Most run the current protocol

About 78% of answering servers speak the current 2025-06-18 MCP spec; the rest trail on older revisions. A long tail frozen on year-old specs is the quiet maintenance signal — a server stuck on an old protocol is usually a server nobody is tending.

Four questions to ask before you connect a server

The numbers above are really four questions worth asking of any single server, in order:

  1. Does it answer? A listing is not a liveness check. Confirm it responds before you build on it.
  2. Does it want your credentials? Know whether a sign-in is required, and to whom you are handing the keys.
  3. What does it cost your context? A 200-tool server is a tax on every turn. Prefer lean, well-scoped ones.
  4. Can you read its tools? The tool names and descriptions are the whole contract. If you can't tell what a tool does, that is the finding.

Method

Serverdex calls every listed server with a remote endpoint using a credential-free initialize + tools/list, about once a day. A server counts as answering when it completes the handshake (with or without a required sign-in); "no MCP answer" means it was unreachable or replied with something that is not MCP. Token sizes are estimated as the characters of the raw tool definitions divided by four — they measure the definitions a client sends, not runtime behaviour.

The numbers here are a snapshot; the live version on Serverdex updates hourly, and the broader ecosystem figures are in The State of MCP Servers.

If you maintain an MCP server: a live endpoint, a readable tool list, and an honest sense of what you're asking people to connect are the whole contract. Write it like one — and know that the people connecting you can, and should, check all three first.

來源:DEV Community · dev.to