✳flâneur — a map of the web's best reading

Skills > MCP | Medium

medium.com · 1,486 words · saved by 1 readers

One GitHub MCP server can expose ninety-plus tools. That’s more than fifty thousand tokens of JSON schemas and descriptions loaded into the model’s working memory before it even starts thinking about your prompt. So when Anthropic introduced Skills, I didn’t see a new feature. I saw a course correction. A quiet one. But a course correction all the same. Under the hood, MCP is pretty simple conceptually. You run one or more MCP servers. Each server exposes a collection of tools, each tool has a schema, and the client loads those definitions into the model’s context so it can decide which tools to call and how to call them. In theory, this gives the model a clean, typed interface to the outside world. In practice, that “clean interface” turns into a firehose. Every tool schema gets stuffed into the context window up front. It doesn’t matter if you’re asking the model to read a file, query a database, or just summarize a paragraph. The model still has to wade through the full inventory of

MCP Is Broken and Anthropic Just Admitted It Cordero Core 6 min read · Nov 13, 2025 -- 68 Listen Share Press enter or click to view image in full size Let’s be honest. MCP works beautifully in demos and breaks the moment you try to scale it. One GitHub MCP server can expose ninety-plus tools. That’s more than fifty thousand tokens of JSON schemas and descriptions loaded into the model’s working memory before it even starts thinking about your prompt. So when Anthropic introduced Skills, I didn’t see a new feature. I saw a course correction. A quiet one. But a course correction all the same. Th

Explore this link on the map →

saved by

related reading