Claude Code MCP servers: just a script, no SDK needed

The distinction between skills and MCP comes from inspecting actual requests. A skill's Markdown body is not sent up front — only its name and description appear in the first request, and the body enters the conversation when Claude invokes it through the Skill tool. An MCP server instead contributes tool definitions and a server-level instructions block near the start of the conversation, which is where you say when this server's tools should be used.

The difference the author considers decisive is who controls what happens after a tool is chosen. With a skill and a script, Claude assembles the command line and can change options, edit the script, or skip it and call the API with curl. With an MCP tool, Claude decides only whether to use it; the JSON Schema and the code constrain the rest. The checklist feature in the author's macOS app tanacode is the example: decisions can only be added, or revoked with a reason, and an attempt to rewrite a card is refused by the app's code. That places MCP between skills, where Claude decides both when and what, and hooks or workflows, where the code decides both.

The concerns behind MCP's name are answered with a log from the memo server showing startup, initialize, tools/list, two tool calls, and the shutdown signal. The author settles on a rule of thumb: skills for procedures and conventions, MCP for operations whose result must be guaranteed — and notes that the two can be combined, putting capability in MCP and long sequencing instructions in a skill to keep the everyday context small.