An MCP server gives an AI agent tools: create an invoice, open a ticket, run a sync. But tools alone don't tell the agent how your product is meant to be used. Which steps come first? What should it check before it does something it can't undo?

The Model Context Protocol now has an official answer. SEP-2640, the Skills extension, has reached Final status, and its PR was merged on September 13, 2026. (MCP: SEP-2640, ext-skills on GitHub)

Is your SaaS agent-ready? Get your score

What a skill is

A skill is a folder of instructions, at minimum one SKILL.md file, that teaches an agent a workflow step by step. It can also include reference files, examples and scripts.

Under SEP-2640, an MCP server can serve those folders alongside its tools. Agents find them with skills/list, fetch one with skills/get, and read the files as ordinary MCP resources at addresses like skill://billing-setup/SKILL.md. The format follows the existing Agent Skills specification, so a skill written once works anywhere that supports it. (MCP Skills overview)

The official docs put the benefit simply: serving skills over MCP "keeps those instructions with the service they describe."

Why this matters if you sell software

Today, the instructions for using your product live in your help center, your onboarding emails and your customer success team's heads. An agent working for your customer sees none of that. It sees a list of tools and guesses.

With Skills over MCP, your product can teach the agent directly:

  • Onboarding becomes a skill. "Set up a workspace, import contacts, connect billing" runs the way you designed it, not the way an agent improvised it.
  • Best practices travel with the product. When you change a workflow, you update the skill on your server and every customer's agent picks it up.
  • Fewer bad first impressions. An agent that misuses your API makes your product look broken. A skill that explains the right order of calls prevents that.

When an agent can't figure out how to use a product well, the customer quietly moves that work somewhere else. That's ghost churn, and skills are a direct fix for it.

One honest caveat

The specification is final, but support in SDKs and AI apps is still rolling out. The MCP project keeps a public client support matrix. (MCP Skills overview) That makes now a good time to plan: build the MCP server and write your first skills while the agents catch up.

Start with the basics

Skills sit on top of an MCP server, so you need a server that agents can discover first. Our free Agent Readiness Check shows whether yours can.

Run the free Agent Readiness Check

Is your SaaS agent-ready? Get your score

Run a free Pryno readiness check on your MCP server