Last updated: 2026-11-20
The short answer
An MCP server gives an agent access to something: a database, an API, a file system. A skill gives an agent know-how: step-by-step instructions, plus optional scripts and reference files, for doing a task well.
MCP answers "what can the agent reach?" Skills answer "how should the agent do this?" Most real workflows use both.
Side by side
| MCP server | Agent skill | |
|---|---|---|
| What it is | A running program that speaks the Model Context Protocol | A folder with a SKILL.md file, plus optional scripts, references, and assets |
| What it adds | Tools, resources, and prompts from an external system | Instructions and procedures for a task |
| Where it runs | As a local process or a remote service | Inside the agent's own environment; nothing to host |
| How the agent loads it | The host connects and lists the server's tools | The agent reads each skill's name and description at startup, and the full instructions only when a task matches |
| Needs credentials | Often, to reach the system behind it | Usually not, unless its scripts call a service |
| Who builds it | A developer writes code against an SDK | Anyone who can write clear instructions in Markdown |
| Defined by | The MCP specification at modelcontextprotocol.io | The Agent Skills format at agentskills.io |
| Typical example | A Postgres server that runs queries | A "write a quarterly report" skill with a template and a checklist |
What an MCP server is
An MCP server wraps a system and describes it in a standard format. The agent's host application connects to it, reads the list of tools it offers, and lets the model call them.
A server is live software. It holds a connection to something outside the agent, and it usually needs credentials. If the server is down, its tools are gone.
Use one when the agent needs to read or change something it cannot otherwise reach.
What an agent skill is
A skill is a directory with a SKILL.md file at its root. The file starts with a short metadata block (a name and a description are required) and continues with instructions in Markdown. The folder can also hold scripts, reference documents, and templates.
Skills load in three stages:
- Discovery: at startup the agent reads only each skill's name and description.
- Activation: when a task matches a description, the agent reads the full
SKILL.md. - Execution: the agent follows the instructions, and opens the bundled files or runs the bundled scripts only if it needs them.
This staged loading is why an agent can have many skills installed without filling its context window.
Use one when the agent can already reach what it needs but does the task inconsistently or badly without guidance.
When to use which
Ask one question: is the gap access, or know-how?
- The agent cannot see your issue tracker → MCP server.
- The agent can see the tracker but writes poor bug reports → skill.
- The agent cannot query the data warehouse → MCP server.
- The agent can query it but ignores your team's metric definitions → skill.
- The agent needs to produce a PDF in your house style → skill, with a template and a script in the folder.
- The agent needs to post the result to Slack → MCP server.
How they work together
A skill can tell the agent which MCP tools to use and in what order. The server supplies the capability and the skill supplies the procedure.
Example: a "weekly incident review" skill. Its instructions say to pull last week's incidents from the monitoring server, group them by service, fill in the review template bundled in the skill folder, and post the summary through the chat server. Two MCP servers do the reaching. One skill makes the result the same every week.
Security differences
Both are third-party content, so read before you install. The risks differ:
- MCP server: the risk is what its tools can do with your credentials. Check the publisher, the tool list, the authentication method, and where the server runs.
- Skill: the risk is what its instructions tell the agent to do and what its scripts run. Read the
SKILL.mdand the file list, especially for skills that include scripts.
A skill with no scripts is plain text that you can audit in a minute. A server is code that you mostly have to trust.
Where to find them
- MCP servers: the SourceWeft MCP directory lists servers with their tools, authentication, and connect guides.
- Skills: the SourceWeft skills directory indexes skills from public GitHub repositories, each pinned to the commit it was scanned at, with the full
SKILL.md, file list, and license.
FAQ
Are agent skills a replacement for MCP? No. Skills carry instructions and MCP carries access. A skill cannot reach a system the agent has no connection to, and a server cannot teach the agent your process.
Can a skill use MCP tools? Yes. A skill's instructions can name the tools to call. The matching MCP server must be connected in the agent's host.
Which one is easier to make? A skill. A minimal skill is one Markdown file with a name, a description, and instructions. An MCP server is a program that you write, run, and maintain.
Do skills only work with Claude? No. Agent Skills is an open format, and a skill can be used by any agent that supports it.
Do I need to host anything to use a skill? No. A skill is a folder that the agent reads. A remote MCP server has to be hosted by someone; a local one runs as a process on your machine.
How do I install a skill from a directory?
In SourceWeft, open the skill and add it to a workspace. For a local agent, run npx @sourceweft/cli skills install <skill-name>.