Last reviewed:
What is an Agent Skill? Definition and difference from MCP
An Agent Skill is a folder containing a SKILL.md file and, where needed, scripts and resources, which teaches an AI agent to perform a task your way. The agent first loads only the name and description, then the detail when the task calls for it. The format, created by Anthropic, is an open standard.
Anthropic introduced Skills on 16 October 2025, then published them as an open standard on 18 December 2025, with a specification on agentskills.io. A skill is a folder built around a SKILL.md file: metadata (a name and a description at minimum) and instructions. It may contain scripts, reference documents and templates. The key principle is progressive disclosure. At startup, the agent loads only the name and description of each skill, about 100 tokens per skill according to Anthropic's documentation. When a task matches the description, it reads the full instructions, under 5,000 tokens. Supporting files are read only if needed, and a script that is run sends only its output into the context. Many skills can therefore stay available without saturating the context window. Anthropic provides ready-made skills for PowerPoint, Excel, Word and PDF. Custom skills are created in Claude Code, uploaded through the API, or added in claude.ai (Pro, Max, Team and Enterprise plans). They do not sync from one surface to another. The agentskills.io site lists many compatible tools, including GitHub Copilot, VS Code, Cursor, Gemini CLI, ChatGPT and Codex. The difference from MCP is one of kind. MCP gives the agent access to tools and data. A skill teaches it a method. According to Anthropic, skills complement MCP servers.
Concrete example
Illustrative case. A 40-person accounting firm writes down its month-end closing method in a skill: the order of checks, the alert thresholds, the format of the summary note and a bank reconciliation script. The accounting software is connected to the assistant through an MCP server, which gives access to the entries. The skill says how to analyse them. A recent hire thus gets a senior's method. Before installation on workstations, the responsible partner reviews the folder, as they would an internal procedure. The firm versions the skill and names an owner.
Comparison
| Prompt | Agent Skill | MCP server | |
|---|---|---|---|
| Role | One-off instruction in a conversation | Reusable method for a task | Access to tools and data |
| Content | Text | Folder: SKILL.md, scripts, documents | Server that exposes functions |
| Loading | Each conversation, by you | On demand, when the task matches | Tools available while the server is connected |
| Main risk | Forgotten instruction, or different from one person to another | Skill from a doubtful source that diverts the agent | Access too broad to sensitive systems |
FAQ
What is an Agent Skill?
It is a folder built around a SKILL.md file, which holds instructions and, where needed, scripts and resources. It teaches an AI agent to perform a specific task your way. The agent loads it on demand, when the task matches its description.
What is the difference between Agent Skills and MCP?
MCP gives the agent access to tools and data (a CRM, a database, files). A skill teaches it a method: how to carry out a task, in what order, in what format. According to Anthropic, skills complement MCP servers. They are often used together.
What is the difference between a skill and a prompt?
A prompt is an instruction given in a conversation, for a one-off task. A skill loads on demand and does not have to be retyped from one conversation to the next: only its name and description take up context until it is used.
Are Agent Skills exclusive to Claude?
No. Anthropic published the format as an open standard on 18 December 2025. The agentskills.io site lists many compatible tools, including GitHub Copilot, VS Code, Cursor, Gemini CLI, ChatGPT and Codex.
Where can I use skills on Claude?
In Claude Code (folders on your machine), in the Claude API (uploaded skills, shared across the workspace) and in claude.ai (Pro, Max, Team and Enterprise plans, individual skills). Custom skills do not sync from one surface to another.
Does a skill carry a security risk?
Yes. Anthropic recommends using only skills from trusted sources, because a malicious skill can steer the agent into misusing its tools or running code, with a risk of data exfiltration. Have the whole folder reviewed before installation.
See also
Further reading
Sources
- Equipping agents for the real world with Agent Skills, Anthropic, 16 October 2025 (updated 18 December 2025). https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills
- Agent Skills, Claude documentation, Anthropic (structure, progressive disclosure, surfaces, security). https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview
- Agent Skills Overview, agentskills.io (open format, compatible clients). https://agentskills.io/home