Tweet by zeeg
January 18, 2026
because im tired of explaining the same thing to every single person who hasnt learned how this all works: 1) mcp is fine. it can use a lot of tokens or a little. its just a framework. github is an example of a bad implementation, sentry is an example of a good. 2) skills use tokens just like mcp does. if you throw 100 skills into your harness you're going to be just as well off as if you threw in 100 random mcp servers 3) lots of the implementations of mcp, skills, commands, subagents, etc share the same characteristics and are roughly experiments to solve the same problem with slightly different ux. context is context. tokens are tokens. 4) there is no magic. theres no magic that lets a harness "load skills" without having the metadata in the system prompt. mcp just made this front and center, and a lot of bad implementations showed up and gave it a bad name. skills will do the same (or already are) 5) as evidence, agent mode in sentry uses about as many tokens (for the entire agent) as one or two well defined skills do
- Author
- zeeg
- Date
- January 18, 2026
- Canonical URL
- /tweets/2026-01-18-zeeg-5306724680-02