A GitHub MCP server burns roughly 50,000 tokens of context just to teach an AI agent how to interact with repositories. A Markdown file that says "use the gh CLI for these operations" does the same job in about 200 tokens. That is a 250x difference, and it is driving a quiet but significant shift in how production AI agent systems are built.
The argument, laid out in a recent piece from The New Stack, is straightforward: most AI agent tasks fall into two buckets. The agent either needs to know something or do something. MCP (Model Context Protocol) was designed for the "do" bucket - live connections to APIs, databases, and external services. But developers kept cramming knowledge into MCP servers too, bloating context windows and burning through API costs for information that never changes.
The Two-Layer Fix
The emerging pattern splits agent capabilities into two layers. Static knowledge - how to format a commit message, what fields a tool page needs, the steps in a deployment workflow - goes into SKILL.md files. These are plain Markdown documents with YAML frontmatter (a small header that tells the agent when the skill is relevant) followed by instructions in regular prose.
Live operations - querying a database, posting to an API, reading real-time data - stay on MCP servers where they belong.
The token savings are not theoretical. That GitHub MCP server's 50,000-token payload gets loaded into the agent's context window (the working memory the AI model can reference during a conversation) every time the agent needs to touch GitHub. A SKILL.md file with the same instructions weighs around 200 tokens. Over hundreds of agent runs per day, that difference compounds into real money.
Who Is Actually Doing This
This is not just a blog post theory. Microsoft shipped a .NET Skills Executor that orchestrates SKILL.md files and treats MCP tools as a subordinate layer underneath. Anthropic opened up its Agent Skills spec as a proposed standard, following the same playbook that turned MCP into an industry-wide protocol. Claude Code already uses this exact pattern: skills folders containing SKILL.md files with reference documents sit alongside MCP server connections.
One particularly striking example: a venture capitalist described running his entire company on twelve Markdown files in a git repo. Structured documents teach Claude Code how to draft emails, triage support tickets, prepare board metrics, and manage product launches. No custom servers. No infrastructure to maintain. Just text files versioned in git.
What This Means for Tool Builders
MCP did not die. It moved down the stack. Skills handle the guidance layer that MCP was never designed for, while MCP handles the plumbing it was built to handle.
For anyone building AI agent workflows today, the practical takeaway is: stop encoding static instructions as tool descriptions in MCP servers. Write a Markdown file instead. It is cheaper to run, easier to debug (you can read it), trivial to version control, and portable across different AI models and frameworks.
The pattern also makes agent behavior more transparent. When an agent misbehaves, you can open a SKILL.md file and read exactly what it was told. Try doing that with a 50,000-token MCP schema dump.
The tools versus knowledge distinction seems obvious in hindsight, but the MCP hype cycle pushed a lot of teams toward using servers for everything. The correction is underway, and the agents are lighter for it.