Agent Skills vs MCP: what a neutral standard means for small business software.

If you build with AI assistants, you have met two terms in the same week: MCP and Agent Skills. Search for agent skills vs mcp and you will find them framed as rivals. They are not. They solve two halves of one problem, and on 22 October 2026 Agent Skills joined MCP at the same neutral home. That is good news for anyone who builds on them, including small businesses that will never write a line of either.

Here is what each one is, how they fit together, what the donation changed, and how we plan to use both.

What MCP is

MCP, the Model Context Protocol, is an open source standard for connecting AI applications to external systems. Its documentation says AI applications like Claude or ChatGPT can use it to reach data sources such as local files and databases, tools such as search engines and calculators, and workflows such as specialized prompts. Its comparison has stuck: MCP is a USB-C port for AI applications. One standard way to plug an assistant into the thing that holds the data.

In practice an MCP server is a small program that exposes a business system to an assistant. Ask about a calendar, a folder or a ledger, and the server is what lets the assistant look. Many assistants and development tools support MCP, so a server built once works in many of them.

What is an agent skill

An Agent Skill answers a different question. Not how the assistant reaches the data, but how it does the job well.

The Agent Skills site calls it a lightweight, open format. A skill is a folder with a file called SKILL.md. That file carries a name, a description and instructions for a specific task. A skill can also bundle scripts, reference material and templates.

The loading is the clever part. At startup the agent reads only each skill's name and description. When a task matches, it reads the full instructions. Then it follows them, opening bundled files only as needed. The site's summary: agents keep many skills on hand with a small context footprint.

Picture it this way. A skill is the written procedure a good employee follows. An MCP server is the key to the filing cabinet. The agent needs both. The foundation's explainer says a skill gives you a place to explain the process, include examples and define how to check the result, while the agent still needs the right access and capabilities to do the task. MCP does not even need to understand what a skill is. To the protocol, a skill is just a resource.

What changed for the Agent Skills standard on 22 October

On 22 October 2026 Anthropic donated Agent Skills to the Agentic AI Foundation, a Linux Foundation project. What moved was the identity of the standard: the name and logo, the GitHub organization and the agentskills.io domain. Anthropic says it is committed to keeping Agent Skills open, community driven and vendor neutral, and that engineers from Anthropic, Google, OpenAI and other companies will maintain it as individuals.

The scale is worth knowing. Anthropic reports that more than 70 agent platforms, including ChatGPT, Cursor, GitHub Copilot and VS Code, have integrated Agent Skills since the standard was published in December 2025. The foundation's blog reports over 1 million skills published and nearly 280 million installations, with the specification open source under Apache 2.0 and CC BY 4.0.

MCP already lives in the same place. Anthropic co-founded the foundation with Block and OpenAI and had donated MCP to it earlier. The two standards an agent most needs, how to connect and how to do the job, now sit under one vendor neutral roof.

For a business choosing what to build on, that matters more than any feature. A format no single company controls is one you can invest in without wondering whose roadmap you are renting.

Agent Skills vs MCP: which one do you need

Our working rule, which is our opinion and not a rule from either standard:

Use an MCP server when the assistant needs to reach something it cannot reach today: your books, your orders, your leads. Use a skill when the assistant can already reach what it needs and the missing piece is the right way to do the work: how your firm reviews a receipt, what your month end checklist is, what a finished report must contain.

Most useful setups use both. The server gives the assistant access, ideally with limited, scoped permissions. The skill gives it the procedure and the check at the end. Because skills are plain folders of text and optional scripts, an owner or an accountant can read one and see exactly what the assistant has been told to do.

What we reviewed and how we apply it

Within three days of the announcement our review panel had screened coverage of the donation. An earlier review, of running an ad platform through an assistant connector, had already produced a plan for SheetTamer.

For SheetTamer the plan has two parts, both on our backlog. First, an MCP server with owner tools, so an owner or her accountant can ask their own assistant about the set aside figure, the receipts to sort and the tax package. Second, downloadable skills shipped beside the server: packaged instructions the owner or accountant loads into their own assistant to read those same three things the way we would. We expect to build them on the open format described above, when the first owner or accountant asks to connect their own tools.

The same idea already shapes how we work. Our build practice keeps its specialist reviewers as written role definitions in the repository, each one pinned to the model that job needs. Writing the procedure down in files anyone can read is the habit the skills format makes portable.

It is also a service. GG Consulting builds AI agents and custom software for clients, which includes writing skills that capture how a team does its work and MCP servers that give an assistant safe, limited access to a client's systems. If you want your process packaged in a format any assistant can use, that is the work we do.

The short version

MCP connects an assistant to your systems. A skill teaches it your procedure. Both now have a neutral home at the Agentic AI Foundation, a base no single vendor owns. Start with one skill for one repeated task, add an MCP server when the assistant needs access it does not have, and keep a person reviewing the result.

Questions about your own setup?

This is a standards story, and your situation may differ. If you are weighing AI tools for your business, we are glad to talk it through.

Contact us