The MCP-vs-CLI debate in AI agent tooling is getting louder, and it's mostly asking the wrong question.
Model Context Protocol (MCP) is the standardized way for AI agents to connect to external services through structured JSON schemas. CLI means letting agents call traditional command-line tools like grep, git, and aws directly. Both work. Neither wins everywhere. The real question is which one fits each specific integration.
Where CLI Still Wins
Command-line tools have a massive advantage that's easy to overlook: LLMs already know how to use them. Models trained on decades of Unix documentation and Stack Overflow answers can call git diff or grep -r without any schema initialization. That's zero startup cost.
The efficiency gap is real, too. CLI pipelines produce roughly 60% fewer tokens than equivalent MCP JSON calls in some workflows. For tight loops, the difference compounds fast. Searching 150 order IDs through MCP means 150 separate tool calls. A CLI for loop does it in about 500 tokens, roughly 1% of the MCP approach.
For local development tools, vendor CLIs with pre-configured authentication, and composable pipelines, CLI is the obvious choice.
Where MCP Pulls Ahead
MCP earns its complexity when you need multi-tenant authentication (multiple users with different OAuth credentials), structured audit logging, or access to services that simply don't have CLIs. Enterprise SaaS integrations are the sweet spot.
The structured schemas also reduce hallucination. When an agent calls a novel API it's never seen in training, having an explicit parameter schema prevents it from guessing wrong field names or formats.
But MCP has a cost problem. A typical GitHub MCP server dumps around 55,000 tokens just to initialize, before any real work happens. At scale, that's roughly $1,600 per day for 10,000 sessions. There's also 50-100ms of serialization overhead per round-trip that adds up in tight loops.
The Missing Layer
The practical solution showing up in tools like Claude Code is a "Skills" abstraction that sits above both protocols. Instead of exposing a full MCP schema dump, a Skill lazy-loads only the relevant parameters for a specific task. A Jira integration might present 2-3 fields at around 300 tokens instead of the full 55,000-token API surface.
This pattern treats MCP and CLI as interchangeable transports underneath a unified interface. The agent doesn't care which protocol is running below. It just calls the Skill.
The takeaway for anyone building agent tooling: stop picking sides. Use CLI for local and composable work, MCP for multi-tenant and stateful integrations, and put a clean abstraction on top. The interface design matters more than the protocol.