Is MCP Dead?

MCP (Model Context Protocol) has drawn attention as a standard for connecting LLMs to external tools, but in actual development workflows, issues around context cost, operational stability, and overlap with existing CLI/API have emerged.

Key Issues

• Excessive context window consumption: When connecting to MCP servers, massive amounts of tokens are consumed just by 'tool definitions' rather than actual task content. According to Quandri's measurements, connecting 4 servers (Linear, Notion, Slack, Postgres) took up 10.5% of Claude's 200K context just for tool definitions. In particular, retrieving a Linear issue via MCP consumes about 65x more tokens than the CLI method.
• Performance and operational overhead: Response speed degrades due to maintaining separate processes, authentication, and round trips to external servers. According to a Jira benchmark, MCP was 3x slower per call compared to direct REST API calls, and 9.4x slower when including initialization.
• Overlap with existing tools: Even though LLMs have already learned how to use CLIs, MCP requires loading separate tool definitions every time, making pipeline composition and debugging difficult.

Alternatives and Improvement Directions

• Claude Code's response: Claude Code recently introduced 'Tool Search with Deferred Loading', which loads tool schemas only when needed, reducing context usage by more than 85%.
• CLI-first and Skills strategy: Instead of the MCP approach of preloading all tools, a strategy using the Skills approach—which selectively loads only the needed functionality—along with existing CLIs is more efficient. This prevents context waste and allows humans and AI to debug using the same interface.