Why the Author Argues MCP Has Become an Outdated Agent Protocol

Summary

This opinion article argues that the Model Context Protocol (MCP), released by Anthropic in November 2024 to connect AI agents with external services and data, was designed for less capable models and has now outlived much of its original usefulness. MCP adoption grew rapidly, but users who added many servers encountered context bloat because every server contributed tools and schemas to the model’s context. Platforms such as Composio, MintMCP and Pipedream responded with generic search-and-execute patterns, centralized credentials and smaller tool surfaces; the author considers these useful short-term remedies, while noting that an ecosystem of monitoring and schema-management systems has also formed around MCP. The article’s central claim is that models have improved enough to execute code, reason over large codebases, compose services and call unfamiliar APIs with limited user intervention. Cloudflare’s Code Mode is cited as an example of using MCP through scripts in a sandbox, while command-line tools can often be discovered through `--help`, eliminating the need for an MCP server when a documented API or CLI already exists. The author therefore recommends removing most MCP servers and standardizing direct agent access to HTTP APIs and CLIs, with mature authentication and content negotiation. Examples include servers honoring `Accept: text/markdown` and documentation sites using `Accept-Language` to return examples for a preferred programming language. The argument is not that every MCP use case has disappeared: the article acknowledges verbose machine-readable CLI output and other remaining problems, but says existing web standards can address them more directly.