OpenRouter analyzes context overhead and token costs for different Agent tools

On November 11, 2026, OpenRouter published "MCP vs CLI vs Skills: What Each Costs in Context", examining how providing different kinds of tools to agents affects context window size and token cost.

The piece starts from a straightforward observation: every tool you give an agent has its definition sent to the model on every request, and those tokens are billed as input. Nothing about that setup is free, and the bill scales with how heavy each tool definition is.

The analysis singles out Model Context Protocol (MCP) servers as the most expensive option, noting that a single MCP server can add tens of thousands of tokens to the context window. Those tokens are consumed on each call, which quickly becomes a meaningful and recurring cost rather than a one-time expense.

For teams looking to reduce the burden, the article points to two more cost-effective alternatives. A command-line tool (CLI) keeps the tool surface deliberately small and lightweight, while a skill offers a structured, reusable way to package capabilities without inflating the prompt. In its framing, CLI and skills deliver comparable functionality to MCP but at a fraction of the context cost.

No follow-up coverage has appeared so far, and no other parties have published figures or counterarguments. OpenRouter's numbers and conclusions therefore stand as the only account of this topic to date.