Skill, MCP Server or Subagent: Which One Costs You Least?
For a chore you repeat, a skill. It is a folder of text files that keeps 49 words in Claude's memory until it runs, with nothing to host.
💵 Cheapest: A skill. Files in a folder, nothing running between uses.
🔌 Connector: An MCP server. Worth it only when Claude must reach a live system.
🧠 Helper: A subagent. A fresh working memory for a big side job, at a bigger token bill.
Skill vs MCP server vs subagent: what is the difference?
All three extend Claude Code. They differ in what runs, and so in what you keep paying for.
| Skill | MCP server | Subagent | |
|---|---|---|---|
| What it is | Written steps (and scripts) in a folder | A program plus a login key | A helper with its own working memory |
| Runs between uses? | No | Yes, with each session | No |
| Best for | A chore you repeat the same way | A live system Claude cannot reach | A big side job |
| Ongoing cost | Your Claude plan only | Keeping the program and key working | More tokens per run |
Think of a kitchen: a skill is a recipe card, an MCP server is an account with a supplier, and a subagent is a second cook in another room. For Tuesday's dinner, you need the card.
Why a skill costs almost nothing while it waits
Claude keeps only a skill's short description in memory. The full steps load when the chore runs.
Our free skill, godmode-lite, waits at 49 words and loads 2,373 when it runs. About 2 percent of it sits in memory the rest of the time.
Which one fits a weekly chore?
Pick one and watch where it stops. Tap a station for the reason.
Figure 1
Weekly chore
Data
Tap a station to see why.
| Step | Skill | MCP | Subagent |
|---|---|---|---|
| Setup | OK: A folder of text files. Nothing to host, nothing left running. | FAIL: A program that starts with every session, plus a login key to the outside service. Worth it only when Claude must reach a live system it cannot reach today. | OK: One text file that describes a helper. |
| Idle | OK: Only its short description waits in memory: 49 words for our free skill. | INFO: Not reached: setup is the cost for a plain chore. | OK: Only its description waits. Nothing runs. |
| Run | OK: Its full steps load only when the chore runs. | INFO: Not reached. | FAIL: It starts with an empty working memory and reads everything again. Anthropic measured multi-agent setups at about 15 times the tokens of a chat. |
| Result | OK: Same written steps every week, so the same result. | INFO: Not reached. | INFO: Not reached. |
Source: Claude Code skills docs (description always in context, full skill loads when invoked); Anthropic engineering, multi-agent research system (about 15x the tokens of chats); godmode-lite SKILL.md 2.5.2 (49-word description).
Core insight: most chores that save you money are written procedures, so they are skills. Add the others only when something real blocks the skill.
When an MCP server or a subagent is worth it
↓
🔌 A live system Claude cannot reach? Add an MCP server
↓
🧠 A job that floods the chat with reading? Hand it to a subagent
↓
✅ Most chores stop at the first line
Subagents are not free. Anthropic measured multi-agent setups at about 15 times the tokens of a plain chat.
Do
- Start with a skill for any chore you repeat
- Add an MCP server when a live system blocks it
Don't
- Run a server for a job a script can do
- Send every small task to a subagent
What does it cost to get one?
You can build a skill yourself for free with our step-by-step skill guide. Or keep the hours: a fixed $250 build, scope written before I start: Get a written scope.
Sources: Claude Code skills docs, code.claude.com/docs/en/skills (description always in context, full skill loads when invoked). Anthropic engineering, "How we built our multi-agent research system" (about 15x the tokens of chats); word counts of godmode-lite SKILL.md 2.5.2, measured 1 October 2026.
Related: what Claude Code skills are, why a skill does not trigger and how many hours a skill saves.