The Model Context Protocol launched in November 2024 as Anthropic's answer to a real problem: LLMs at the time couldn't reliably interact with external services. MCP gave them a standardized schema layer — tool definitions, structured inputs, predictable outputs. For models that couldn't write a cURL command, this was genuinely necessary plumbing. The argument against MCP in 2025 is not that it was wrong, but that it was scaffolding. The author, returning from a full-day MCP conference, makes a structural claim: models can now read --help output, compose multi-service scripts, and call undocumented APIs with minimal human intervention. Cloudflare's Code Mode, launched September 2025, makes the case explicitly — it uses LLMs to compose MCP tool calls into executable scripts, effectively routing around the protocol's per-call overhead. The middleware became the bottleneck. Context bloat is the presenting symptom. Each MCP server ships its own tool schemas, and agents loading dozens of servers burn context window on definitions rather than work. A secondary industry emerged — Composio, MintMCP, Pipedream — offering generic search/execute wrappers that reduce the schema footprint. These are good short-term patches, but the author's point is that patching a protocol whose core assumption (models can't call APIs) is now false is architectural debt, not progress. The alternative stack is already deployed. HTTP APIs with standard authentication, content negotiation via Accept headers, and CLI discovery are mature, widely documented, and require no proprietary intermediary. The author highlights two concrete adoptions: documentation sites honoring Accept: text/markdown to serve agent-optimized responses, and Shopify implementing Accept-Language headers so agents receive SDK-specific docs. These are protocol-level improvements to the existing web, not a parallel ecosystem. MCP was donated to the Agentic AI Foundation under the Linux Foundation in December 2025, a move that reads as institutional preservation of a protocol whose original sponsor may already see diminishing returns. The MCP ecosystem now includes monitoring platforms, schema managers, credential brokers — a classic middleware tax that grows its own constituency regardless of whether the underlying problem still exists. The honest complication is that not all services have well-documented CLIs or APIs. MCP still provides value for integrations where the alternative is screen-scraping or reverse-engineering proprietary interfaces. The author's "delete most of your MCP servers" is more precise than "delete all." The question isn't whether MCP dies, but whether the ecosystem built around it can accept graceful deprecation or will lobby for its own continuation. What makes this argument worth tracking is the velocity of the underlying shift. The gap between "models can't call APIs" and "models read --help and compose scripts" closed in roughly eight months. Protocol decisions made for November 2024 capabilities are being stress-tested by July 2025 capabilities. The meta-lesson: any protocol designed around a capability ceiling will be obsoleted by the next capability jump, and AI capability jumps are arriving faster than protocol governance cycles can adapt.
