🤔 MCP - Evil
Morgan Linton (CTO of Perplexity) believes that MCP is not justified, and I completely agree.
MCP is a lot of wasted context and tokens, unmanageable and glitchy. Some MCP servers (like GitHub) load over 40k tokens just for tool descriptions - and all of this hangs in the model's context, even if it's never used.
Modern agents, with documentation, can write and run code on the fly, call APIs directly, and manage perfectly without MCP. CLI tools, direct access to the file system, code - all of this works more reliably and efficiently. Anthropic is moving in this direction with their Agent API.
On Twitter, there is a contrary opinion - that MCP provides standardization "write once, support every AI client", which is convenient for enterprise and stateful interactions. But in practice, this is more of a nice idea than a working solution.
What do you think?
