On September 13, 2026, the Model Context Protocol (MCP) added the MCP skills extension via Standard Enhancement Proposal (SEP) 2640 [1]. A skill is a folder of text files that teaches an Artificial Intelligence (AI) agent to do a task [1]. Any MCP client can now find these files at a skill:// URI (web address) using two new commands: skills/list and skills/get [2].

Splitting the Agent Stack

Why it matters: The AI software stack is splitting in two. MCP handles how an agent talks to tools, while skills handle how an agent learns a process [3].

Before this shift, builders loaded servers and skills in separate steps [1]. Strict size limits also meant builders could not pack large workflows with their tools [1]. SEP-2640 fixes this by setting one clear way to send and grab skills over the network [3].

By the numbers:

  • 875 lines: The size of the mcpGraph skill. This proves the need to bypass old server limits [1].
  • 3 methods: The setup uses skills/list, skills/get, and an optional resources/directory/read to browse files [2].
  • 512 files or 16 MiB: The strict size limit per skill. Servers must stay under 16 Mebibytes (MiB) [2].
  • 2 layers: The standard cleanly splits the software stack into tools (MCP) and processes (Skills) [3].

How to Adopt MCP Skills

The playbook: AI system builders should make four key moves to use this standard:

  • Treat skills as resources: A skill is a group of files under a skill:// address. Systems that view MCP resources as a virtual file system can read skills just like local folders [1].
  • Tie approvals to exact files: Every entry has a file list with Secure Hash Algorithm 256-bit (SHA-256) digital fingerprints and exact byte sizes. Tie your security approvals to the full file set. Any changed file instantly breaks access [2].
  • Do not trust skill content: Require clear, per-skill user consent before running code or granting allowed-tools access. Reading files across different servers also requires fresh consent for every call that names both servers [2].
  • Wire the fleet: Use new fleet tools to deploy servers. For example, Claude Code 2.1.259 uses managedMcpServers to push Hypertext Transfer Protocol (HTTP) and Server-Sent Events (SSE) network servers across a company [5]. It also adds a --permission-prompts none flag to lock down quiet agents [5].

The Security Burden Remains

The catch: This new rule only sets how systems move skills, not how they format them. Folder layouts and YAML (a data format) tags still rely on the separate Agent Skills guide [1][4]. MCP and Agent Skills remain two distinct rule sets that must stay perfectly in sync.

The security burden also lands on hosts. The creators removed the archive format because opening a remote server’s zip file invites nasty attacks [1]. Skill content remains untrusted input. This forces a boring but highly secure setup of sending files over resources.

  • The Big Shift: The Model Context Protocol (MCP) officially added the MCP skills extension on September 13. This turns folders of instructions into simple network resources that systems can easily find and read.
  • Why It Matters: The Artificial Intelligence (AI) agent stack is splitting into two clear layers: tools (MCP) and processes (Skills). This turns skills into safe, verified network resources instead of messy custom folders.
  • The Winning Moves:
  • Treat skills as resources: Read files under the skill:// address as a virtual file system for an easy setup.
  • Tie approvals to exact files: Link security approvals to Secure Hash Algorithm (SHA) fingerprints so any file tweak revokes access.
  • Do not trust skill content: Demand clear, per-skill user consent before running code or granting tool access.
  • Wire the fleet: Deploy Hypertext Transfer Protocol (HTTP) servers securely across a company using fleet tools like managedMcpServers.
  • The Catch: The extension only sets rules for moving skills, not formatting them. Hosts must also handle the security risks of untrusted inputs since zip files are banned to stop attacks.

Related reading

Sources

[1] SEP-2640: Skills Extension — Model Context Protocol — https://modelcontextprotocol.io/seps/2640-skills-extension [2] Skills — Model Context Protocol — https://modelcontextprotocol.io/extensions/skills/overview [3] ext-skills repository — Skills Over MCP Working Group — https://github.com/modelcontextprotocol/ext-skills [4] Agent Skills specification — https://agentskills.io/specification [5] Claude Code changelog — https://code.claude.com/docs/en/changelog