MCP Was a Mistake for Local Development. Here's why CLI + modern agent harnesses win on tokens, speed, and reliability — and the progressive path to fixing it.
A fresh agent session with several MCP servers enabled can burn 50,000–100,000 tokens on tool schemas before you type a single instruction. Those tokens cost money and reasoning capacity. For local tools the model already knows — git, docker, jq, kubectl — that schema tax is almost pure waste.
This video is a progressive critique, not a blanket rejection: - The visible token tax of loading schemas - Why models are already fluent in CLI (and why MCP throws that fluency away) - The composability loss when you can't pipe output - Where MCP still genuinely belongs (remote SaaS, multi-user auth, sandboxes without a shell) - The 4-step progressive path: switch to CLI, build agent-friendly CLIs, harness them as tools, design for the hybrid future
If you're running local MCP servers for git, databases, or build tools — this video shows you exactly what to replace and how.
TIMESTAMPS 0:000:00 The 50,000-token tax 1:401:40 Why MCP got applied beyond its sweet spot 2:272:27 The familiarity mismatch: models already know CLI 3:143:14 Composability: one command vs many round-trips 4:014:01 The deeper architectural problem 4:284:28 Where MCP still belongs 5:125:12 Progressive solution 1: stop the bleeding 5:455:45 Solution 2: build agent-friendly CLIs 6:146:14 Solution 3: harness the CLI as a tool 6:486:48 Solution 4: the hybrid future 8:208:20 What you should do next