MCP is dead?

This post on the Quandri engineering blog continues the argument from "MCP is dead. Long live the CLI," using the company's own real-world development environment to take measurements. The author argues that the core problem with MCP (Model Context Protocol) is not that the concept is wrong, but that it often becomes too heavy in everyday engineering workflows. The article points out that after connecting to MCP servers such as Linear, Notion, Slack, and Postgres, the tool definitions themselves consume a large amount of context; Quandri measured that 4 servers with 77 tool definitions amounted to roughly 84,308 characters and about 21,077 tokens, occupying about 10.5% of Claude's 200K context and about 16.5% of GPT-4o's 128K context. Linear alone has 42 tools and about 12,807 tokens, so even if a user only needs to look up or update an issue, they still bear the cost of the entire schema. An update to the article also adds that Claude Code subsequently introduced deferred-loading Tool Search, which can substantially reduce context bloat, but the author believes that concerns regarding performance, debugging, and architecture still remain.