Skills over MCP - Ola Hungerford, Principal Engineer Nordstrom

Ola Hungerford, principal engineer at Nordstrom and an MCP maintainer, explains the new Skills extension for the Model Context Protocol and how it went from an open question to an accepted spec change. This talk is for MCP server authors who want to ship agent skills alongside their tools, and for anyone who wants to know how a change actually gets into MCP. Tool descriptions tell an agent what each tool does, but not how to combine them for a task, what order to use them in, or what conventions your domain follows. That is what agent skills are for. Until now, skills and MCP servers traveled through two separate channels. The Skills extension serves skills from the same server as the tools: one channel, one set of access controls, one way to version and distribute both. - Why skills and MCP complement each other: tools are the raw materials, and skills are the specialized knowledge and preferences you do not want the agent to rediscover every time. - The two-channel problem: skills installed from a plugin marketplace, MCP servers configured separately, and two different ways of controlling access. - Shipping the manual in the same box as the tools: the idea behind serving skills over MCP. - Options the working group considered: a new skills primitive, server instructions, bare MCP resources, and dedicated skill-listing tools. - Why they landed on MCP resources, and the two parts of the extension: listing skills on a server and reading them progressively. - Three read steps: the skill list with URIs and frontmatter passed through unchanged, then SKILL.md on activation, then supporting files. - Skill identity defined as the server plus the skill's resource URI. - A digest for every skill and every file, so clients can tell whether an approved skill has changed. - A new optional method for reading a directory, since MCP resources had no equivalent to a directory listing. - From interest group to working group to SEP-2640, and what the iteration felt like from the inside. - Lessons from the review: keep it minimal, avoid too much optionality, and remember it is easier to add things later than take them away. - What is still open: efficient search and loading across many tools, resources, and skills, progressive disclosure in MCP, and a file systems working group. If you maintain an MCP server and already publish skills somewhere else, this is the extension to adopt next. LINKS AND RESOURCES - SEP-2640, the Skills extension proposal: https://github.com/modelcontextprotocol/modelcontextprotocol/pull/2640 - The official Skills extension repository: https://github.com/modelcontextprotocol/ext-skills - Skills Over MCP Working Group charter: https://modelcontextprotocol.io/community/working-groups/skills-over-mcp - Model Context Protocol: https://modelcontextprotocol.io/ - Agentic AI Foundation: https://aaif.io/ TIMESTAMPS 00:00 Introduction: Ola Hungerford, Nordstrom and MCP maintainer 01:25 How agents, tools, and skills fit together 02:27 When to let the agent explore and when to give it a skill 03:41 The two-channel problem for skills and MCP servers 03:59 Shipping the manual in the same box as the tools 04:38 Options considered: a new primitive and server instructions 05:00 Why bare resources and specialized tools fell short 05:20 Where they landed: MCP resources 06:00 The three read steps 06:10 Passing agent skills frontmatter through unchanged 07:06 Resolving skill identity 07:32 Digests for every skill and file 07:56 An optional directory read method 08:20 From interest group to working group 09:10 SEP-2640 and getting to final 10:28 Lessons: add later rather than take away 10:56 Why too much optionality causes fragmentation 11:16 Who showed up to build it 11:39 What the extension does not solve yet 13:04 How to get involved in MCP #ModelContextProtocol #AgentSkills #SkillsOverMCP