MCP Is Being Abandoned: How Fast Can a 'Standard' Die?

In mid-March, Perplexity's CTO Denis Yarats casually dropped a bombshell at the Ask 2026 conference: the company is moving away from MCP internally, going back to REST APIs and CLIs. The audience barely reacted, but the statement exploded on social media. YC CEO Garry Tan retweeted it with a blunt "MCP sucks honestly" — it eats too much context window, authentication is broken, and he wrote a CLI wrapper in 30 minutes to replace it. A year ago, this kind of pushback would have been unthinkable. MCP was hailed as the ultimate standard for AI tool integration, ecosystem growth was explosive, and server counts doubled weekly. Now it's been hyped, overused, and rejected. So what actually went wrong with MCP? MCP Is Too Heavy: Context Windows Can't Handle It A standard MCP setup consumes roughly 72% of the context window. Someone measured it: three servers (GitHub, Playwright, and an IDE integration) burned through 143K tokens of tool definitions on a 200K-token model. Before the agent even