Skills vs MCP: one teaches the agent how, the other connects it to what
MCP (Model Context Protocol) is an open protocol that connects an AI agent to external tools and data, while an Agent Skill is a folder of instructions, scripts and resources that teaches an agent how to do a specific task and loads only when the task calls for it.
Checked
MCP and Agent Skills both started at Anthropic, and both are now open standards, so they are often compared as alternatives. They solve different problems. MCP answers how an agent reaches a system. A skill answers how an agent should do a piece of work once it can reach it.
Anthropic's own framing is that MCP supplies connectivity and skills supply expertise. Most production agents end up with both.
What each one is
MCP
An open protocol for exposing tools and data to an agent through MCP servers. It is one of the founding projects of the Linux Foundation's Agentic AI Foundation, alongside AGENTS.md and Block's goose.
Agent Skills
A folder with a SKILL.md file plus optional scripts and resources. Only the name and description sit in context until the agent decides the skill is relevant, then the rest loads. Anthropic calls this progressive disclosure.
Portability
Skills became an open standard in December 2025, and OpenAI's Codex documents support for Agent Skills built on the same standard.
How they combine
Anthropic recommends one general agent specialised with a library of skills, rather than a separate agent for each domain. In that design, MCP servers give the agent access to systems such as the CRM, the helpdesk or the data warehouse. Skills tell it how to run a month-end reconciliation or triage a support queue using those systems.
A skill can also carry deterministic code. Anthropic's guidance on code execution with MCP suggests saving working code that calls MCP tools as a reusable skill.
The context cost
Loading every MCP tool definition up front, and passing every intermediate result through the model, spends context the agent needs for the task. Anthropic reports that presenting MCP servers as code APIs, so the agent loads tools on demand and filters data in a sandbox, took one example from about 150,000 tokens to about 2,000.
Skills are cheap by design, because only a short description is preloaded. The practical rule: keep MCP tool sets small and focused, and move procedure and judgement into skills.
When to reach for which
Build or adopt an MCP server
When the agent needs to read or change a system it cannot reach yet, and more than one agent or client will need that access.
Write a skill
When the agent can already reach the systems but does the task inconsistently: a house style, a checklist, a review procedure or a multi-step runbook.
Use both
When a repeatable task spans several systems. Motiontic's engineers usually ship the MCP connections and the skills library together, each with its own evals.
MCP compared with Agent Skills
| MCP | Agent Skills | |
|---|---|---|
| What it carries | Access to tools and data | Instructions, scripts and resources for a task |
| Question it answers | What can the agent reach? | How should the agent do this work? |
| Format | A server that speaks the protocol | A folder with a SKILL.md file |
| Loading | Tool definitions, often loaded up front | Name and description only, the rest on demand |
| Context cost | Grows with every tool exposed | Small until the skill is used |
| Standard | Founding project of the Agentic AI Foundation | Open standard since December 2025 |
Common questions
Do Agent Skills replace MCP?
No. MCP connects an agent to tools and data. Agent Skills teach the agent how to carry out a task. Anthropic describes MCP as supplying connectivity and skills as supplying expertise, and most production agents use both.
Can a skill use MCP tools?
Yes. A skill can tell the agent which MCP tools to call and in what order, and it can include code that calls them. Anthropic suggests saving working tool-calling code as a reusable skill.
Which uses more context, MCP or skills?
MCP usually does, because tool definitions are often loaded up front and intermediate results pass through the model. Skills preload only a name and description, and the full instructions load when the skill is used.
Do Agent Skills only work with Claude?
No. Agent Skills became an open standard in December 2025, and OpenAI's Codex documentation describes support for skills built on that standard.
When should a company build its own MCP server?
When agents need access to an internal system that has no existing server, and more than one agent or client will use that access. For a single agent and a single system, a direct tool can be simpler.
Related
MCP vs A2A
The protocol for tools and data, compared with the protocol for agents talking to agents.
Context Engineering vs Prompt Engineering
Wording the instructions, compared with curating every token the model sees.
Agent harness
The loop, tools, context, permissions and state that sit around the model.
AGENTS.md
A README for coding agents: the file an agent reads first in a repo.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.