Model Context Protocol (MCP) and agent skills are complementary. An MCP server provides live data, authentication, authorization, and actions. A skill tells the agent when and how to use those capabilities to complete a workflow.
Separate access from procedure
| Concern | MCP server | Agent skill |
|---|---|---|
| Live data | Retrieves current bookmarks | Explains which bookmarks to select |
| Permissions | Enforces authorized operations | Sets workflow limits and approval points |
| Tool calls | Exposes available actions | Defines sequence and decision logic |
| Output | Returns data or action results | Defines the final format and quality bar |
| Testing | Checks tool contracts and authorization | Checks activation and workflow behavior |
Use MCP without a skill
Use an MCP server by itself when the user controls each action and the task does not need a stable procedure. Searching recent bookmarks for a phrase is an access task. The server can return matching records without a separate skill.
Use a skill without MCP
Use a skill by itself when the required information is already in the prompt or packaged references. Reviewing an exported bookmark file requires a procedure but does not require live account access.
Use MCP and a skill together
Combine them when the workflow needs current bookmarks and consistent judgment. For example, an MCP server can retrieve bookmarks through read-only access. A skill can then qualify candidates, inspect linked sources, reject unsupported claims, and produce a review report.
The combined workflow follows this sequence:
- Retrieve the authorized bookmark set through the MCP server.
- Apply the skill's qualification criteria.
- Use approved research tools for missing context.
- Produce the output defined by the skill.
- Request approval before any controlled write or installation.
Keep the security boundary explicit
Do not use skill instructions as a substitute for server-side authorization. The MCP server must enforce access controls. The skill must explain when to call the tools, which results to trust, and when to stop.
OpenAI documents this same boundary in its official skills and MCP guidance. To decide whether you need a workflow at all, compare an X bookmark manager with an agent skill.