The Context Overload Problem and Code Mode

▶ Watch (2:28)

Early MCP clients loaded every tool’s metadata into context at session start. Each tool consumed about 500 tokens. A single GitHub MCP server with 100 tools burned 60,000 tokens. Two or three such servers filled the context window. Code mode, released by Cloudflare on September 26, 2024, solved this by exposing tools as code APIs. The agent searches a small API, reads documentation, then runs scripts in a sandbox. Only two or three tools remain in context, costing roughly 2,000 tokens. Drawbacks: it requires sandbox access, kills interactivity, and can lower task completion if the search tool is not optimized.

Skills: Reusable Prompts with Progressive Disclosure

▶ Watch (8:53)

Anthropic introduced skills in October 2025. They are markdown files containing prompts, optional scripts, assets, and references to sub-skills. The front-end design skill, the third most downloaded, has over 200,000 downloads. It is a 30-line prompt. Skills load only the name and description at session start. The full file enters context only when the user asks for it. Sub-skills load only when needed. At Alpic, a skill builds MCP apps autonomously: brainstorming, coding, testing, and deploying via the Alpic CLI. The entire process takes 30 minutes. Skills provide instructions, not tools. They work best inside coding harnesses that have terminal access.

CLI: The Old Standard

▶ Watch (16:06)

CLIs are deeply embedded in model training data. They offer progressive discovery through --help, are token-efficient after 50 years of optimization, and compose via Unix piping. Drawbacks: non-technical users cannot use them. They lack standard authentication, multi-tenancy, audit trails, and analytics. Service providers have no control over usage. CLIs remain the best choice for local development workflows and terminal-friendly users.

When to Use What: A Mental Model

▶ Watch (17:23)

MCP alone works for simple workflows and UI interactions. Example: ordering groceries on the Instacart MCP app in ChatGPT. Skills alone are reusable prompts for writing content and code. Skills plus MCP handle complex workflows with external services. Alpic uses a skill plus MCP to enrich inbound contacts and push them to CRM. Skills plus API call GraphQL endpoints, like Alpic’s Linear skill. Skills plus CLI automate local PR reviews. The combinations cover everything from non-technical users to developers.

Q&A

Is MCP going to die? Rodionov said no; MCP is healthier than ever because it is a standard accessible to anyone, unlike CLI. ▶ Watch (20:20)

Will MCP become MCPS (skills integrated into the protocol)? Rodionov agreed they are complementary and converging. A working group in the MCP protocol is already developing skills over MCP. ▶ Watch (20:52)

Notable Quotes

“MCP has been announced dead many many times.” Nikolay Rodionov · ▶ Watch (0:28)

“saying thank God MCP is dead just as useless of an idea as LLM.text was.” Nikolay Rodionov · ▶ Watch (1:48)

“Code mode is basically just an optimization of how you call the MCP tools.” Nikolay Rodionov · ▶ Watch (7:55)

“MCP tools are connectors that expose new capabilities to your models, and skills are recipes that tell your model how to use these capabilities.” Nikolay Rodionov · ▶ Watch (15:22)

“It’s not dead. Is it going to die? I don’t think so. It’s never been healthier than today.” Nikolay Rodionov · ▶ Watch (20:20)

Key Takeaways

  • Code mode reduces context from 60,000 tokens to 2,000 by exposing tools as code APIs.
  • Skills are reusable prompts with progressive disclosure; they do not replace MCP.
  • MCP, skills, and CLI are complementary and already converging in practice.

About the Speaker

Nikolay Rodionov is a cofounder and COO at Alpic, an AI infrastructure startup for MCP, and creator of Skybridge, an open-source framework to build MCP Apps. He previously founded Streamroot, a P2P video delivery technology that powered multiple Super Bowl streaming events before…