Scoped Skills Mcp And Memory — Claude Code Guide

Tools, skills, MCP, and memory are different layers

Tools describe actions, skills describe method, MCP provides external sources, and memory stores facts or agreements that may persist between runs. Do not enable every layer for every agent by default.

Least privilege applies to all four: a narrow boundary reduces excess context, false assumptions, and unwanted actions.

Component	Purpose	Question before enabling
Tools	Perform actions	Which tool is indispensable?
Scoped skill	Define workflow and checklist	Which method fits this role?
MCP	Obtain external data	Are local files sufficient? Is write needed?
Memory	Store durable facts	Is this a lasting rule rather than a hypothesis?

Scoped skills and MCP

A reviewer may use a diff-risk analysis skill.
A tester may use a test strategy and regression skill.
A migration assistant may use read-only documentation lookup.
Do not connect a database or issue tracker “just in case”.
Allow write access through MCP only for a separately justified operation.

Memory must be durable

Store stable naming conventions, test locations, and verified team rules. Do not store one-off hypotheses, temporary stack traces, flaky-test assumptions, secrets, or tokens.

A fresh reviewer may be more useful without memory so outdated rules cannot distort its conclusion.

Short tester memory example

unit tests: backend/src/test/java
refund regressions: use the Refund prefix

Do not store:
- one-off hypotheses;
- temporary stack traces;
- credentials or tokens.

Enable components in order

Define one clear role.
Grant the minimum tools.
Add only the relevant skill.
Check whether MCP is needed and leave it disabled if local sources suffice.
Decide whether memory is needed.
Ask whether the agent can perform the role without each component.