# Why MCP is dead & How I vibe now Source: https://www.youtube.com/watch?v=fG95XsBO5U4 Recap page: https://rapidrecap.app/video/fG95XsBO5U4 Generated: 2026-01-24T10:33:37.746+00:00 --- ## Quick Overview The presentation argues that the Model Context Protocol (MCP) is a dead concept, advocating instead for a shift towards a more token-efficient and scalable architecture based on Skills combined with a Command Line Interface (CLI), exemplified by Vercel's agent-browser which achieved a 70% reduction in token consumption compared to traditional Chrome Dev Tool MCP methods. **Key Points:** - Vercel's agent-browser demonstrated a 70% less token consumption in tests compared to the Chrome Dev Tool MCP implementation. - The author strongly asserts that the MCP concept is dead, favoring a Skill + CLI approach as being significantly better and more scalable. - The Chrome MCP tools consumed approximately 4,000 tokens for an instruction, whereas the agent-browser CLI instruction in CLAUDE.md only required 95 tokens. - The presentation introduces a hierarchical action space for context offloading, consisting of Function Calling, Sandbox Utilities, and Packages & APIs, where each layer balances capability, cost, and cache stability. - The shift from MCPs to Skills allows for greater agent capability expansion (hundreds of integrations) without bloating the context window with unnecessary tool definitions. - The speaker demonstrates creating a new custom skill for MCProxy using the `add-new-mcp` skill, showing how easily new MCP servers can be integrated and documented via Skills. ![Screenshot at 00:05: The presenter shows a comparison slide highlighting that Vercel's agent-browser achieves "-70% less token consumption compared with Chrome dev tool MCP in my test", illustrating the core argument for adopting Skills + CLI over MCP.](https://ss.rapidrecap.app/screens/fG95XsBO5U4/00-00-05.jpg) **Context:** This video serves as a technical argument and demonstration advocating for a paradigm shift in AI agent architecture, moving away from the Model Context Protocol (MCP) structure towards a more efficient system based on specialized 'Skills' integrated via a Command Line Interface (CLI). The speaker uses the example of Vercel's 'agent-browser' to illustrate the massive token savings achieved by this new approach compared to older MCP methods, positioning the Skills architecture as the future for scalable agent development. ## Detailed Analysis The video argues that the Model Context Protocol (MCP) approach is obsolete and advocates for replacing it with a Skill + CLI architecture, which is significantly more token-efficient and scalable. The speaker cites a test where Vercel's agent-browser (using the new approach) resulted in 70% less token consumption than the Chrome Dev Tool MCP, noting a specific example where 4k tokens were reduced to just 95 tokens when using the CLI instruction via the agent-browser skill. The proposed architecture utilizes a Hierarchical Action Space with three levels: Function Calling (standard, but cache-breaking), Sandbox Utilities (allowing shell execution like `bash` for large outputs), and Packages & APIs. The speaker then demonstrates the process of creating a new skill (`add-new-mcp`) to easily integrate a new MCP server (like the Excel MCP) by defining its schema in a Skill file, showing that the agent can then directly call these specialized functions, saving context space that would otherwise be consumed by loading dozens of MCP tools. ### MCP vs. Skill/CLI Argument - MCP is a dead concept - Skill + CLI is far better and more scalable - Agent-browser showed 70% token reduction vs. Chrome MCP (4k tokens vs 95 tokens) ### Context Offloading - Three abstraction levels: Function Calling (cache breaks easily) - Sandbox Utilities (model calls shell utilities like bash) - Packages & APIs - Each layer balances capability, cost, and cache stability ### Skill Implementation Details - Skills load only necessary context (e.g., Context7 skill only cost 135 tokens) - MCP tools cost 48.7k tokens in comparison (for Claude-Haiku context) ### Creating Custom Skills - Demonstrated adding an 'excel' MCP skill by creating a SKILL.md file detailing workbook operations (create, get metadata, copy, rename, delete) - Demonstrated adding a 'superdesign' skill via `npx skills add` command, which auto-generates necessary files and documentation ### Superdesign Demo - Used the 'superdesign' skill to iterate on a design prompt, fetching guidelines via the 'context7' skill, demonstrating complex, multi-step agent workflows that are context-efficient. ![Screenshot at 00:05: The presenter shows a comparison slide highlighting that Vercel's agent-browser achieves "-70% less token consumption compared with Chrome dev tool MCP in my test", illustrating the core argument for adopting Skills + CLI over MCP.](https://ss.rapidrecap.app/screens/fG95XsBO5U4/00-00-05.jpg) ![Screenshot at 00:25: A context usage breakdown showing that MCP tools consumed 48.7k tokens \(24.4%\) in the context window, visually emphasizing the token overhead being discussed.](https://ss.rapidrecap.app/screens/fG95XsBO5U4/00-00-25.jpg) ![Screenshot at 00:48: A slide summarizing the Context Offloading approach with three abstraction levels: Function Calling, Sandbox Utilities, and Packages & APIs, emphasizing balancing capability, cost, and cache stability.](https://ss.rapidrecap.app/screens/fG95XsBO5U4/00-00-48.jpg) ![Screenshot at 03:19: A slide detailing Level 3 Resources and code \(loaded as needed\) for a Skill, showing that instructions, code, and resources \(like utility scripts\) are bundled directly within the skill structure.](https://ss.rapidrecap.app/screens/fG95XsBO5U4/00-03-19.jpg) ![Screenshot at 08:50: The terminal shows the successful installation of the 'superdesign-skill' via npx skills add superdesigndev/superdesign-skill, followed by the agent recognizing the skill and its available commands.](https://ss.rapidrecap.app/screens/fG95XsBO5U4/00-08-50.jpg)