What is an MCP and how does it work?

MCP (Model Context Protocol) is Anthropic's standard protocol for connecting AI models to external tools and services. A Supabase MCP, for example, allows Claude to read and write to your database directly from the session, without you having to manually copy and paste schemas or data.

How many MCPs should I install?

The Level 4 principle is surgical selection. For most projects, 2–4 well-chosen MCPs are sufficient. More than 6–8 active tools starts to degrade Claude's ability to choose the right one. Install what you need for the current project, not what you might need someday.

How do I know if Claude chose the right tool?

Ask it: «Why did you use that tool and not this one?» If the explanation makes sense, good. If it can't explain it or the answer is vague, it's a sign that the selection was heuristic, not reasoned. At Level 4 you develop the habit of asking for explanations after every non-obvious decision.

Are frameworks like BMAD or GSD necessary?

They're not necessary, and in many cases native Claude Code produces better results. Frameworks have value for teams that want a predefined working structure. The Level 4 recommendation is to experiment with a framework, understand what it adds and what it limits, and make an informed decision about whether to adopt it for your specific context.

Level 4 — The Toolsmith

13 March 2026

Mindset: «You equip Claude»

At Levels 1–3, Claude works with what you give it in the chat. At Level 4, you connect Claude to the real world: it can read your database, test your UI, push to GitHub, query external APIs. You stop working with code in a vacuum and start building real products.

The key mindset is that you are the one who equips Claude — not the other way around. You choose the right tools for the right job, project by project. Surgical selection always beats accumulation.

You're here when...

You install MCPs (Supabase, Figma, Playwright, GitHub...)
You connect Claude to the outside world and external services
You experiment with frameworks like BMAD or GSD
You actively explore what capabilities each tool adds
You start building end-to-end flows, not just code snippets

Skills to master

Surgical MCP selection — right tool, right job, per project (not all tools all the time)
Understanding building blocks: frontend / backend / auth / databases / security / deployment
Asking Claude «Explain HOW you did that» after every non-obvious decision
Using Claude as an infinitely patient tutor for concepts you don't understand
Knowing when native Claude Code outperforms a third-party framework

Trap: The candy shop

Capability does not equal performance. 15 tools loaded = Claude picks the wrong one. More options = worse decisions. «Just one more MCP, bro» is the classic Level 4 trap.

Native implementations increasingly outperform layered frameworks. If you're dealing with user data, auth, or client projects, you must understand the conceptual layer of what's happening. «Claude told me to do it» is not a valid answer if something goes wrong.

How to unlock Level 5

The unlock comes when you notice you're repeating the same workflows manually over and over. Project setup, code review, deployment, documentation — sequences that could be encoded once and run automatically. That's Level 5: skills.

Next steps

If you want to design the optimal tool stack for your team and your projects — without falling into the accumulation trap — we can help you select it.

Frequently asked questions

What is an MCP and how does it work?: MCP (Model Context Protocol) is Anthropic's standard protocol for connecting AI models to external tools and services. A Supabase MCP, for example, allows Claude to read and write to your database directly from the session, without you having to manually copy and paste schemas or data.
How many MCPs should I install?: The Level 4 principle is surgical selection. For most projects, 2–4 well-chosen MCPs are sufficient. More than 6–8 active tools starts to degrade Claude's ability to choose the right one. Install what you need for the current project, not what you might need someday.
How do I know if Claude chose the right tool?: Ask it: «Why did you use that tool and not this one?» If the explanation makes sense, good. If it can't explain it or the answer is vague, it's a sign that the selection was heuristic, not reasoned. At Level 4 you develop the habit of asking for explanations after every non-obvious decision.
Are frameworks like BMAD or GSD necessary?: They're not necessary, and in many cases native Claude Code produces better results. Frameworks have value for teams that want a predefined working structure. The Level 4 recommendation is to experiment with a framework, understand what it adds and what it limits, and make an informed decision about whether to adopt it for your specific context.