MCP, MCP 2.0, SKILLS, and Sub-agents
MCP, Skills, and Sub-agents solve three different problems. MCP connects Claude to things outside itself. A Skill tells Claude how you want a job done. A Sub-agent hands the job to a separate instance that works on its own.
MCP (Model Context Protocol)
MCP is a protocol for connecting a model to tools, databases, and data sources it never saw in training.
Without it, Claude works from training data with a cutoff date plus whatever you paste into the conversation. With it, Claude reaches:
- Live data such as weather and stock prices
- Company systems such as a customer or product database
- Tools such as code interpreters, calculators, and mail systems
- Files on your own machine
How a request flows
- You run an MCP server that fronts a tool or data source
- Claude sends it a request: fetch this, run that
- The server does the work and returns the result
- Claude uses the result in its answer
MCP 2.0
MCP 2.0 revises the original design rather than replacing it. The headline change is a stateless architecture: each request carries everything the server needs, so the server keeps no session for you between calls.
Compare it to asking a librarian for a book. Under MCP 1.0 the librarian tracked your whole visit history to answer. Under 2.0 they answer the question in front of them.
Dropping per-client state buys three things:
- Speed. No session to look up or maintain.
- Reliability. One failed request cannot corrupt the next.
- Simpler servers. Less bookkeeping to write and get wrong.
Build new servers on 2.0. Existing 1.0 servers keep working, so migrate them when you have a reason.
Skills
A Skill is a set of written instructions for a task you repeat. It lives in a file and gives Claude the steps you want, in the order you want them.
A Skill holds:
- Instructions in markdown
- A description that says when to use it
- The steps to follow
A skill called write-tech-news might say: search for this week’s tech stories, confirm each one in a second source, draft the post, commit it, push to GitHub. Ask for the Skill and Claude runs those steps.
Skill against MCP
MCP gives Claude reach: a database it can query, an API it can call. A Skill gives Claude a procedure: the steps you want followed. One is a new capability, the other is a job description.
Sub-agents
A Sub-agent is a separate Claude instance scoped to one kind of work. You start it with its own instructions, it works through the task on its own context, and it reports back.
Common ones:
- A code reviewer that reads diffs and flags bugs
- A researcher that searches and summarizes
- A deploy agent that ships to servers
- An explorer that maps an unfamiliar codebase
Sub-agent against Skill
A Skill runs inside your conversation and spends your context. A Sub-agent runs beside it with a context of its own and hands back a result. A Skill says “here are the steps, follow them.” A Sub-agent says “go handle this and report when it’s done.”
All four side by side
| MCP | MCP 2.0 | Skill | Sub-agent | |
|---|---|---|---|---|
| What it is | Protocol for outside connections | Stateless revision of MCP | Written workflow | Separate Claude instance |
| Gives you | Access to tools and data | The same, with less overhead | Repeatable steps | Focused work off your context |
| Cost | Depends on the tool | Lower than 1.0 | Cheap | A second instance to run |
| Reach for it when | You need live data | You are writing a new server | You repeat a process | A task deserves its own context |
Picking one
Reach for MCP when Claude needs something it cannot see: a live feed, a database, a file on disk.
Reach for a Skill when you find yourself typing the same five instructions every week.
Reach for a Sub-agent when a task would flood your main conversation, or when two pieces of work can run at once.
Using them together
A weekly tech-news post uses all three. MCP 2.0 pulls the articles. The Skill holds the steps for turning them into a post. A Sub-agent runs the whole job while your main session does something else.