We treat AI agents as first-class readers of our documentation. The same Taiga UI knowledge is published in a few complementary shapes — a flat context file, a query server, and reusable skills — so whatever your assistant can consume, it reaches the current, version-correct API instead of guessing from memory.
That matters more than it sounds. An assistant can scaffold a screen in seconds, but it generates components from training data frozen in time — so the code often compiles yet quietly targets an API that has moved on, and a growing app drifts out of sync with itself. Point the agent at a live, versioned source and generated code stays aligned with the components you actually ship.
These aren't competing options — they are a stack of roles over one body of knowledge.
llms-full.txt is the generated data, @taiga-ui/mcp is a query interface over that same data, and skills are the method that consumes it. When they disagree, the live source wins.