MCP vs CLI vs Skills: When a Shell Command Is Enough and When You Need a Server

What happened

The article analyzes token costs and technical trade-offs between CLI, MCP, and skills in AI agent workflows. It presents data from the author's servers, benchmarks like the 32x token gap, and explains how on-demand loading optimizes context usage.

Why it matters

Understanding these trade-offs helps developers optimize AI agent performance, reduce context costs, and manage stateful operations effectively in tool integration.

• Prioritize CLI for familiar, stateless commands to minimize token costs.
• Use MCP for tools requiring state, credentials, or structured outputs.
• Leverage skills for contextual guidance without preloading large datasets.
• Implement on-demand loading to avoid unnecessary token consumption.
• Measure tool costs in your specific workflow before choosing a protocol.

The bigger picture

The article influences decisions about tool architecture, context management, and cost optimization for AI agents, particularly in environments requiring stateful operations or credential security.

The article provides a technical framework for choosing between CLI, MCP, and skills, emphasizing context costs, tool requirements, and on-demand loading strategies for AI agent development.

Technical details

• MCP server schemas are loaded per request (7,553 tokens total for 14 tools).
• CLI help texts are loaded only when needed, reducing per-turn costs.
• Skills use declarative rules for activation and can dynamically load MCP servers.
• On-demand capability activation limits active tools to 4 at a time.
• Token counts use OpenAI's o200k_base tokenizer (Claude's tokenizer would vary).

FAQ