Quickstart

In about fifteen minutes you can ask an AI assistant to teach you a topic from sources you choose, answer the first question of a lesson built for you, and come back days later to review what you are about to forget. This page works from a checkout of the repository, which also gives you the demo lessons. To connect an assistant with Node only and no checkout, see Add the server to your assistant. Each step links to the page with the details.

Self-learner journey: a request in an AI assistant, references and their licences, exact passages, an agent-drafted lesson, an outline review, play with the authoring notice visible, then reviews when concepts fall due.

What you need

Item Why
Node 22 (20 at least) and pnpm 9.15.9 build and run the Studium server
Git clone the repository
An AI assistant that speaks MCP: Codex, Claude Code, Copilot in VS Code, OpenCode or Cursor it relays the lesson to you
VS Code (optional) the lesson pane with figures, equations and reviews

No model API key is needed by Studium itself. Your assistant keeps its own account and terms. MCP (Model Context Protocol) is the standard way an assistant calls outside tools; see the glossary.

1. Install and build

git clone https://github.com/HolobiomicsLab/studium.git
cd studium
./setup.sh

setup.sh installs the workspace, builds the libraries and the VS Code extension, and checks the result. It produces apps/studium-vscode/dist/lesson-mcp.cjs, the Studium server your assistant will start. Details and the manual equivalent: Installation.

2. Connect your assistant

Register the server once. With Claude Code, from the checkout:

claude mcp add --scope project studium -e LITELLM_BASE_URL= -- node "$PWD/apps/studium-vscode/dist/lesson-mcp.cjs"
claude mcp list

With Codex:

codex mcp add studium --env LITELLM_BASE_URL= -- node "$PWD/apps/studium-vscode/dist/lesson-mcp.cjs"
codex mcp list

LITELLM_BASE_URL= (empty) tells Studium to use no judge model, even if your shell defines one; step 3 relies on it. Copilot, OpenCode and Cursor have their own pages, listed in Connect your assistant; the direct HTTP API is in the HTTP API reference. Without further settings the server keeps your lessons in ~/.studium/library and your progress in ~/.studium/data, so it survives a restart.

3. Ask to learn something

Open a new conversation in your assistant and write, in your own words:

I want to learn how a hash table handles collisions with Studium.

Clients that support MCP prompts also offer the server's learn prompt. The assistant then:

  1. reads your study plan (get_study_plan), which says what to do next;
  2. asks you for sources (a web page, a PDF, notes) if no lesson covers the topic yet;
  3. builds a lesson from them (get_authoring_guide, prepare_sources, create_lesson): every quote is checked character for character against your sources before the lesson is written;
  4. starts it (start_lesson_session) and shows the first question.

The first screen says the lesson was drafted by an agent and not checked by a teacher. Answer in your own words; the assistant sends your text (submit_response) and shows the feedback. Without a judge model, open answers are not scored: you get the reference to compare with yourself. The full walk-through is the tutorial Learn a topic in a conversation.

4. Come back to review

A few days later, ask:

What should I review today with Studium?

The assistant calls get_study_plan, then start_review_session to play every review that is due, across your lessons. When you rate how well you recalled an answer (again, hard, good, easy), only the next review date moves; your measured mastery does not. See Review over time.

Prefer a visual pane

To play the same lessons in VS Code, install the extension (see VS Code), then run Studium: Pick a course or right-click a lesson.json and choose Studium: Open this lesson. Your assistant can also hand a lesson over with a vscode:// link.

Next steps

You are Go to
A self-learner Learn with your assistant
A student given a course Student guide
A teacher Teacher guide
Stuck Troubleshooting

View the source of this page on GitHub