Skills or MCP: Which Do You Need?

You have an agent running, and it works. Now you want to give it something it does not have yet. You go looking, and you get back a pile of answers that all sound like they compete for the same job: Skills, tools, MCP, just let it write code. Not one of them tells you when you would pick it over the others. The question I get asked more than almost anything else about agents is this one. Skills or MCP, which should I use? It has a clean answer, and it is that those two are not competing at all. Think about a new person starting on Monday. They need a badge, because without one they cannot get into the building. And they need to know how your team does the job: where the runbook lives, what gets checked before a release, who signs it off. The badge does not teach them the job, and the handbook does not open any doors. You have never once looked at a new colleague and asked yourself, badge or handbook. MCP is the badge. Skills are the handbook. Once you can see which question each one answers, the choice gets easier, and you can also see when to reach for something else instead. Covered: The vocabulary, said plainly: agent, tool, context window, token, and what it means for an agent to run code rather than write it for you to read The axis every option is really chosen on: how much of your context window does this eat, and WHEN does it eat it All four ways to give an agent capability side by side, with what each costs you and who decides to use it Why a tool definition is a standing cost, paid on every turn whether it is called or not Progressive disclosure, the piece I would keep from the whole episode: three levels loading at three different times, and the ratio that breaks the link between how much capability you have and what it costs per turn The detail almost nobody knows: when the agent RUNS a script from a skill, the code never enters the context window, only its output does. And the condition that reverses it Why a skill with a perfect body and a vague description may never get opened at all, and what changes when skills live in the same commit as the code they describe What you accept when you install someone else's skill, and the review-path gap most teams have Where MCP sits after the donation to the Agentic AI Foundation, and how A2A is different By the end there is a four-line card on screen that routes the decision, and that card is your answer if somebody asks you this in an interview. Sources and dating Every claim comes from published material, dated on screen rather than in the narration. A best-effort snapshot as of 8 September 2026: guidance and APIs change quickly, so check the docs and the cookbook before relying on any figure or API detail. Ten of the eleven sources are Anthropic's; the eleventh, on A2A, is the Agentic AI Foundation's. Equipping agents for the real world with Agent Skills (Anthropic Engineering, 16 Oct 2025) Building agents with the Claude Agent SDK (Anthropic, 29 Sep 2025) Code execution with MCP: building more efficient agents (Anthropic Engineering, 4 Nov 2025) Donating the Model Context Protocol, establishing the AAIF (Anthropic, 9 Dec 2025) Agent Skills documentation, Skill authoring best practices (platform.claude.com, 7 Sep 2026) Skills notebooks, CMA_use_skills_from_a_repo (`anthropics/claude-cookbooks`, 3 Sep 2026) A2A joins AAIF's open agentic stack (Agentic AI Foundation, 17 Aug 2026) Two things are flagged on screen as practical advice rather than documented guidance: writing a "do not use for" boundary into a description, and the invoice figures in the run-versus-read example. 00:00 Skills or MCP, and why it reads like a menu 01:13 A new person starts on Monday. Badge, or handbook? 02:45 The words, said plainly: agent, tool, context window, token 04:20 A filesystem, and an agent that runs code itself 05:13 MCP is the badge: one protocol, not a connector per pairing 06:08 Skills are the handbook: a folder, and procedural knowledge 07:38 The better question: which of the two are you short of? 08:22 Four ways to give an agent capability, and the two axes 09:18 1 of 4, tools: a standing cost, paid on every turn 11:11 2 of 4, bash and scripts: you pay only for what comes back 12:07 3 of 4, code generation, and why a bundled script is different 13:41 4 of 4, MCP, and the four side by side 14:58 Progressive disclosure, and the three levels 16:23 The script's code never enters context. And when it does 17:32 What the numbers mean, and the ratio that transfers 18:38 The description is a trigger, not documentation 20:44 Two skills fighting over one request 21:52 Composability: one skill pointing at another 23:16 Skills in the repo, in the same commit as the code 26:07 A skill is code, and what you accept when you install one 28:14 MCP after the donation, and how A2A is different 30:22 A gap I could not close, said plainly 30:54 The card, and back to Monday morning