Why MCP Has Outlived Its Purpose as Agents Learn to Call APIs Directly

The Model Context Protocol, which Anthropic shipped in November 2024 and handed off to the Linux Foundation’s Agentic AI Foundation a year later, solved a real problem for its era: early LLMs weren’t capable enough to reliably orchestrate external tools on their own, so a standardized connector layer gave agents dependable access to outside services and data. Adoption exploded alongside the broader LLM boom. But that success bred its own overhead — stacking multiple MCP servers floods a model’s context with tool schemas, spawning a cottage industry of workarounds from vendors like Composio, MintMCP, and Pipedream that centralize credentials and prune tool exposure to keep context lean.

The author’s core argument is that the models simply outgrew the protocol. Today’s agents can execute code, reason over large codebases, and compose unfamiliar APIs into working scripts with little hand-holding. They discover CLIs by running —help and hit documented HTTP endpoints directly — the very APIs most remote MCP servers were merely wrapping. Cloudflare’s Code Mode, which has LLMs assemble MCP calls into sandboxed scripts, is cited as a signal that even MCP’s proponents are routing around its limitations.

The proposed path forward is to delete most MCP servers and lean on existing infrastructure: documented APIs, standard content negotiation, and mature auth. Rather than inventing new protocol machinery, agents should use what the web already offers — content-negotiation headers like Accept: text/markdown to get clean text instead of bloated HTML or verbose JSON, or Accept-Language-style hints so docs sites return SDK-specific examples (an idea Shopify’s Tobi Lütke agreed to ship). The remaining friction, mainly token-heavy machine-readable CLI output, is treated as a solvable engineering detail rather than a reason to keep MCP alive.