Curriculum is the product.
Inference is yours.

Most AI-assisted courses ship as a hosted product: you log into a web app, it talks to a model on your behalf, and the price reflects both the curriculum and every token spent answering you. The instructor picks the model, the prompt, the safety rails — and bills you for compute they cannot meaningfully control.

You already pay for an agent. Claude Code, Cursor, Codex, Zed — pick your seat. Paying again for inference baked into a course price is double-charging for the same capability.

A course is two things glued together: curriculum (the structured progression of ideas, exercises, grading) and an agent (the thing you talk to, the thing that runs the code). The first is hard to author and worth paying for. The second has become a commodity you already own.

ngguide separates the two. The curriculum lives on a server you connect to. Your own agent does the talking. You pay once for the course, never for inference.

Model Context Protocol gives agents a way to call tools on remote servers. That's the perfect shape for pedagogy: the server holds state your agent shouldn't (progress, mastery, the next lesson, the right answer), and exposes a small vocabulary of tools — get_current_lesson, submit_answer, hint, mastery_check.

The agent stays the agent. The course stays the course. Neither has to pretend to be the other.

Your agent forgets between sessions. The server doesn't. Progress, mastery, spaced-repetition schedules, deterministic grading — all of it lives on the curriculum side, where it can be authored once and trusted everywhere.

That also means switching agents mid-course is free. Start in Claude Code, finish in Cursor, review on Zed. The state follows the curriculum, not the client.

You pick the agent. You pick the IDE. You pick the model. You read the code, write the answers, run the tests. The course is your tutor — not your stand-in. We don't have an opinion on which seat you bring, only that you bring one.

If a better agent ships next month, you can adopt it without us shipping anything.

Inference will keep commoditizing. Agents will keep getting better. The lasting craft is curriculum: sequencing ideas, picking exercises that build skill, grading sharply enough that students cannot bluff past mastery.

That's the part we want to be excellent at. Everything else — model choice, IDE chrome, autocomplete behaviour — belongs to the student.

Cohort features, peer review, auto-generated lessons, certificates, agents that complete the course on your behalf. None of these are part of v0.1, and several may never be.

The pilot is intentionally small: one course, one author, one curriculum surface. We will widen it once we know which constraints matter and which were just defaults.