Skills vs MCP

A skill is a markdown file, sometimes a folder of them, that an AI coding agent reads as part of its working context. The file tells the agent when to apply itself and what to do when it does.

Agent skills use an open format. The skills.sh registry indexes them across agents and provides one command that installs any of them.

How agent skills work

Without Stream Agent Skills installed, an agent has to guess how to work with Stream based on whatever it learned during training and whatever it can find by searching online. With the skills installed, the agent reads our authoritative SKILL.md before writing any code and pulls live doc references when it needs them.

This means with the skills you know that the SDK versions in your generated code are current, specific function usage is correct, token generation follows our recommended patterns, and more - generally, you'll see fewer errors.

How skills differ from MCP

MCP (Model Context Protocol) is a runtime protocol that connects an agent to tools and data sources through a long-lived server process.

A skill, by contrast, is static markdown that ships once and is read on demand. There's no server to run and no connection or state to maintain.

A skill is read on demand and then it's done. An MCP server has to start up and hold a connection open the whole time the agent might reach for it:

flowchart TB
    subgraph Skill["Agent skill"]
      direction TB
      S1["Install once<br/>markdown lands on disk"] --> S2["Agent reads SKILL.md<br/>on demand"] --> S3["Generates code"]
    end
    subgraph MCP["MCP server"]
      direction TB
      M1["Start server process"] --> M2["Agent connects,<br/>connection stays open"] --> M3["Agent calls tools<br/>over the connection"]
    end

    style Skill fill:transparent,stroke:transparent
    style MCP fill:transparent,stroke:transparent

Project info

Stream Agent Skills are released under the Apache 2.0 license. Bug reports and pull requests are welcome at GetStream/agent-skills on GitHub.