MCPs & Sub-agents

MCPs teach Claude how to interact with external services. You can ask Claude to find and recommend MCPs instead of searching yourself.

You don't need to manually search for MCPs — just ask Claude "find me a good Figma MCP." Be aware that Claude's knowledge might be slightly outdated, so verify the recommendations. MCPs are powerful but come with the token cost trade-off.

Don't manually configure MCP settings files. Ask Claude to install and configure MCPs for you.

No more editing config files manually like you had to with Cursor. Claude can handle the entire MCP installation and configuration process. Just say "install this MCP for me" and approve the actions.

Use sub-agents for atomic, parallelizable tasks — not for work that needs the main context. Only the output returns, not the reasoning.

Common mistake: using sub-agents for tasks that need the full conversation context (like a "testing agent" that needs to know about code changes). Sub-agents only return their output summary, not their full reasoning chain. Use them for isolated, atomic tasks (architecture audits, code scanning) and keep context-dependent work in the main session.

Keep your context condensed and fresh. Don't overload with sub-agents, MCPs, and massive rule files simultaneously.

The "second brain" approach: lazy-load context, keep rules concise, use sub-agents sparingly. Build a personal RAG system of project knowledge that you pull in as needed rather than loading everything upfront.