Risky Business Features MCP is Dead
Mar 18, 2026
A contrarian take that the Model Context Protocol has lost relevance and why that matters for AI tooling. A look at how models are skipping structured tool protocols in favor of direct shell and CLI interactions. A rundown of the security tradeoffs as agents move from managed MCP servers to raw shell access. Practical advice on where to focus access controls and auditability.
AI Snips
Chapters
Transcript
Episode notes
MCP Was The Tooling Breakthrough For LLMs
- MCP unlocked LLMs by packaging capabilities as tools the model could call to act in virtual and physical systems.
- Anthropic's MCP let models list tools, authenticate, and choose when to call them, creating powerful agent behavior.
Claude Code Showed MCP In Practice
- Claude Code exemplified MCP power by shipping curated tools for creating, editing, diffing, and running code from a thin terminal UI.
- Teams rapidly built MCP servers locally then in the cloud, adding OAuth/API keys to scope access.
Wholesale API To MCP Mapping Caused Context Pollution
- Many MCP servers mirrored entire existing APIs, creating huge tool lists that polluted model context and forced many atomic tool calls.
- That wholesale mapping made MCP interactions token-inefficient and brittle versus curated high-leverage tools.