Agent Skills - a thin alternative to the Model Context Protocol?
Just a year passed since the Model Context Protocol (MCP) has been introduced by Anthropic and despite of the traction it got in 2025, it got already concurrency by Anthropic itself: Agent Skills.
Also, it seems that OpenAI is experimenting with the idea.
This makes it worth to have a short look at the differences between the two. I also post some links below if you want to see examples for Agent Skills.
While the Model Context Protocol is a whole Specification which needs to be implemented or used via libraries, Skills are simple Markdown files which describe how an LLM can access what to reach a goal. Anthropic has an example in their blog for generating a PDF where the pdf/SKILL.md contains python code to execute.
Having just a markdown file with additional scripts is much thinner than the implementation of a whole Protocol.
Compared to the Model Context Protocol skills need access to the filesystem, CLI and tools for execution. This has some security implications, whereas the MCP has more control about authorization and access control and implements the code to be executed by itself.
In my opinion both can exists side-by-side: Skills are good for local usage to show models how it can assist with daily work, whereas MCP is nice for companies which want to provide access to their services and have more control over the execution path.
Also, in a company context, an MCP server can be setup and provided to a whole bunch of internal/external users and controlled from one central place. This can be done for users, which are not technically competent and they don’t want to have all the pain of experimenting with models.
But I think skills can lead to faster results, as they are so thin and don’t need the downloading or implementation steps which MCP servers have. The files can even be written by an LLM itself to iterate fast about different descriptions.
Security is a topic for both sides although in my opinion the security considerations for Skills also apply for MCP and the other way around: Always check external sources, what models can access and get instructed to do!
The future will show which one survives and gets a higher adoption. I’m sure we’ll also see a third way to enhance model capabilities emerging.
Resources
- Anthropic Agent skills - a guide to skills
- Anthropic skills repository - examples for skills
- Codex docs skills.md - how Codex describes skills
- Claude Skills - simonwillison.net - a nice overview and comparison between Skills and MCP.