Perplexity ditched MCP. 72% of the context window was consumed by tool schemas before the agent did anything.
Cloudflare replaced MCP with Code Mode, code generation against a typed SDK. 2,500 API endpoints that would need ~244,000 tokens as MCP tools? ~1,000 tokens. A 244x reduction.
Scalekit benchmarked five identical GitHub tasks:
-
CLI: 4-32x fewer tokens than MCP
-
Reliability: CLI 100%, MCP 72%
-
Cost per 10K operations: CLI $3.20, MCP $55.20
Garry Tan posted "MCP sucks honestly" and shipped a CLI-based agent setup instead. 30+ CVEs filed against MCP in 60 days. 82% of surveyed servers vulnerable to path traversal.
72% of the context consumed before the agent starts. 72% reliability when it does.
I keep seeing the same pattern: LLMs were trained on man pages, READMEs, CLI documentation. Small tools, text streams, composability. The Unix philosophy is exactly what agents understand best.
MCP isn't dead. It solves real problems for tool discovery and enterprise multi-tenant scenarios. But for developer tools and local agents, the command line was always the answer.
Where are you drawing the line between CLI and MCP in your agent stack?