MCP server or skill? The question that decides most setups

A server adds capability; a skill adds a rule. Most confusion comes from asking a skill to do something only a server can, or the reverse.

Nov 20, 2026The avbbc desk
MCP server or skill? The question that decides most setups
A server adds capability; a skill adds a rule. Most confusion comes from asking a skill to do something only a server can, or the reverse.

Two extension mechanisms, one question: does the agent need to be able to do something new, or to do something the same way every time? A server answers the first; a skill answers the second. Nearly every confused setup is a mismatch between the two.

This page is the decision rule, with examples from both directories.

The short answer

If the agent cannot reach the data or perform the action at all, you need a server. If it can, but does it inconsistently, you need a skill. A skill that needs new capability fails; a server that encodes a rule is overkill.

The four cases

The agent… Needs Example
cannot read the data MCP server PostgreSQL
can read it, uses it inconsistently Skill SQL Analyst
cannot perform the action MCP server Vercel deploy read
can act, needs a consistent process Skill Release Notes for support

The second and fourth rows are the ones people try to solve with a server, which produces a bigger integration that still behaves inconsistently.

Why the confusion exists

Both are installed the same way, both change behaviour, and both are described as 'giving the agent new abilities' in marketing. The difference only shows up when something breaks: a server failure is an error, a skill failure is wrong output.

The failure modes

A skill asked to do a server's job produces confident fabrication: the agent has no data, so it writes plausible content. This is the most damaging mismatch in the directory.

A server asked to do a skill's job produces a large integration with no discipline: the agent can query anything and queries the wrong thing. The fix is a small skill on top, not a bigger server.

The architecture that works

One server per source, read-only by default. One skill per repeated judgement, with a check. The skill uses the server; the server does not know about the skill. That layering is why a read-only server plus a good skill beats a bespoke integration every time.

A worked example

A team wants an agent that answers product questions from a database. The server is Postgres, read-only. The skill is SQL Analyst: one query, explicit columns, a row limit, and an explanation that names the table. The server provides access; the skill provides the discipline. Neither alone produces a trustworthy answer.

Frequently asked questions

Can a skill call a server? Yes, that is the intended composition. The skill states when to use which tool.

What if I only have time for one? Start with the server if the data is unreachable, the skill if the output is inconsistent. See MCP in practice.

Are prompts a third category? A prompt is an instruction for one task; a skill is a standing rule. See prompts.

Related reading

All posts