A blog post titled "Why MCP Was Always a Bad Idea" argues the Model Context Protocol has been overtaken by the same large language models it was built to serve. Developer Maharshi Patel's post drew more than 170 points and 128 comments on Hacker News.
Patel wrote that MCP made sense when Anthropic introduced it in 2024, giving models a structured way to discover and call external tools. He argued that today's models write scripts, call undocumented APIs directly and reason about large systems well enough that the protocol's schema and tool-discovery layer now add overhead rather than remove it. Loading several MCP servers into one session, each carrying its own tools and schemas, can overwhelm a model's context window, he wrote, a problem platforms like Composio and MintMCP have built workarounds for.
As an alternative, Patel pointed to Cloudflare's Code Mode approach, which has models write and run sandboxed scripts that compose API calls directly, and to emerging conventions such as an "Accept: text/markdown" header, already adopted by Shopify's documentation, that let a model request content in a format built for it instead of a person.
Whether MCP is obsolete or just overdue for a narrower job is exactly the argument now running through those 128 comments. Anyone maintaining an MCP server for an internal tool has likely felt the tradeoff Patel describes: a protocol built to make tool use legible to a model becomes one more thing to maintain once the model gets good enough to skip it.