Data and analysis on the MCP server ecosystem, from the index behind MCP Rating.
Our "UI-capable" badge came from keywords and package descriptions. Its most-downloaded hit was the SDK itself, and it missed a server whose MCP Apps views were already in our own database. So we measured 79 servers instead.
Read →In September a local 31B tried four servers from our index and none started. Part of the reason was us. Our search buried the servers people use, and our gateway recommended servers it couldn't start. After fixing both, the same models connect on the first try, every time. What still fails is API keys, and one hosted server asked a model for an email address. The model made one up.
Read →We started measuring MCP clients and found that what a client announces in its handshake tells you very little. Then we found the same gap in our own API, our own site, and — eight times — in the instruments we were measuring with.
Read →Standing overhead is not what a conversation costs. Measured end to end with a real tokenizer, an MCP gateway costs 2.7× more than a static config at one server, breaks even at three to five, and reaches about 3× at ten — not the 10× we were quoting.
Read →We pointed three local models at a gateway to 63,000 MCP servers. From 8B up they discover a server, connect to it, and call a tool that did not exist when the conversation started. Asked for a web search server specifically, the best of them tried four and none of them started.
Read →Counting MCP servers is easy and everyone does it. Starting them is not, and nobody publishes it. Here is what happened when we connected to 182 and asked what they expose.
Read →Counted six ways, the MCP ecosystem ranges from 2,225 to 371,199 servers — a 167x spread that depends entirely on a definition nobody states.
Read →