Agent skill vs MCP server
A factual comparison of agent skills and MCP servers, what each supplies, and when a workflow needs one, the other, or both.
- Published
- Sep 27, 2026
- Checked
- Sep 27, 2026
- Words
- 897
How far we checked this page
Signed by Arun Kumar, Sep 27, 2026. Passing one check says nothing about the next.
- FormatNot checkedThe file format was not checked for this page.
- InstallsNot checkedInstallation was not checked for this page.
- LoadsNot checkedLive loading was not checked for this page.
- HelpsNot checkedA with/without task comparison was not run.
An agent skill is a folder of instructions that teaches a coding agent how to carry out a repeatable workflow. An MCP server is a live connection that exposes tools, resources or prompts from an external system: it supplies capability that plain instructions cannot, such as authenticated access to real data. Reach for a skill when the job is a procedure the agent’s existing tools can already handle. Reach for an MCP server when the job needs an external system the agent cannot otherwise reach. As of September 2026, many real workflows use both: a skill that says what to do, backed by a server that gives it something to do it with.
Choose an agent skill when
- The workflow is a repeatable procedure (a checklist, a review process, a house style), and the tools the agent already has are enough to carry it out.
- You want something portable: a
SKILL.mdfile that follows the Agent Skills specification is plain text, and travels between hosts more easily than a server connection does. - You don’t need live external data or an authenticated action, only consistent instructions.
Choose an MCP server when
- The task needs something a skill’s instructions can’t supply on their own: live data, an authenticated API, or a callable action on a system outside the agent’s workspace.
- The same capability needs to be available across many tasks and skills, not just one procedure.
- You’re prepared for the operational side of running or trusting a server. MCP’s own security guidance flags authorization and local-server risks that a plain skill doesn’t carry.
Worked example
Say the job is a weekly issue-triage summary. A skill named triage-issues could instruct the agent to group open issues by recurring failure signature, cite each issue’s link, and draft a short summary in a fixed format. If the issues already sit in an exported file inside the workspace, that skill can do the whole job with nothing but the agent’s normal file-reading tools, with no MCP server involved.
The moment the task needs the current state of a live tracker (issues filed an hour ago, or the ability to close one once it’s triaged), the skill’s instructions alone can’t get there. A GitHub MCP connection supplies that: it exposes tools to list, read and act on issues, and the skill can instruct the agent to call them. OpenAI’s skill-and-MCP guidance draws this same boundary: the server supplies data, authentication and actions; the skill supplies the workflow.
We haven’t built or run this pairing ourselves; it’s a design pattern from the docs, not a tested integration. If you build something like it, the two things worth checking by hand are whether the server actually connects (a written config isn’t proof of a live connection) and whether the skill’s instructions actually call the tool rather than guessing at an answer.
At a glance
| Agent skill | MCP server | |
|---|---|---|
| What it is | A folder of instructions (SKILL.md plus optional scripts/, references/, assets/) | A process or endpoint exposing tools, resources or prompts |
| Defined by | Agent Skills specification | MCP specification, revision 2026-07-28 |
| Supplies | Know-how: steps, checklists, house style | Access: live data, authenticated actions, external systems |
| Discovery | Metadata (name, description) loads first; instructions load on activation | Tools listed via tools/list, called via tools/call |
| Executes code? | Not by itself; it can instruct the host to use its existing tools, and only runs its own scripts/ if the instructions call for them | Not necessarily; many servers do trigger actions or code, but a read-only server can serve only data, resources or prompts |
| Portability | The SKILL.md format is portable; host-specific fields and folder paths aren’t | Depends on host support for MCP and the server’s own transport |
| Needs credentials? | Not by itself | Often; most useful servers sit behind auth |
What we checked
Everything above the worked example is drawn from published specifications and docs, not from our own testing of an MCP connection: the Agent Skills specification, the MCP specification and OpenAI’s skills-and-MCP explanation, all checked September 27, 2026.
The one thing we did test hands-on, on the same date, was skill loading on its own, not paired with MCP. We put a small skill (explain-error, with instructions and a references/checklist.md) at .agents/skills/explain-error/ in a scratch Codex project, and confirmed on Codex CLI 0.157.1 that it was discovered, that a matching prompt triggered it with both files read (visible in the tool trace), and that an unrelated prompt didn’t. That speaks to the skill side of this comparison in general; it says nothing about MCP, which we did not install or connect for this page.
What we didn’t check: we didn’t connect an MCP server of any kind, so we can’t confirm connection behavior, tool-call latency, or context cost for MCP on any host. We also didn’t test skill loading in Claude Code (the CLI couldn’t authenticate non-interactively in our test setup) or Cursor (not installed). Treat the Codex loading result above as version-specific until someone runs the equivalent elsewhere.