Choose your model
Studium ships no model and picks none. This page says where a model is used, what it must be able to do, and how to work with a Copilot plan, with a model that runs on your machine, or with no extra model at all.
Before you start
- Connect an assistant first: Connect your assistant. MCP is the protocol an assistant uses to call Studium's tools.
- A judge is a model that assesses a written answer; see Honesty and provenance.
Where a model is used
| Model | What it does | Where you choose it | Needed |
|---|---|---|---|
| Your assistant's model | builds a lesson from your sources, shows each step, passes your answers on | in the assistant (Copilot, Claude Code, Codex, OpenCode, Cursor) | yes |
| A judge model | assesses the audit steps, where you say what is wrong in an AI's reading or summary, and leads the discussion steps | on the Studium server, with LITELLM_BASE_URL |
no |
The two are independent. A Copilot plan gives your assistant its models; it does not give Studium an endpoint, so it cannot serve as the judge.
Your assistant's model
Studium asks one thing of it: the model must be able to call tools, since every step of a lesson is a tool call.
With a Copilot plan
Add the server to Copilot in VS Code:
code --add-mcp '{"name":"studium","command":"npx","args":["-y","studium-mcp"],"env":{"LITELLM_BASE_URL":""}}'Open Copilot Chat in agent mode and pick a model in the model picker. The models offered depend on your plan (Pro, or Copilot for students and teachers through GitHub Education): see Connect Copilot.
Ask: "I want to learn … with Studium", and name your sources.
With a model that runs on your machine
| Assistant | How |
|---|---|
| VS Code chat | run Chat: Manage Language Models, then Add Models: the Ollama extension for a model served by Ollama, or Custom Endpoint for any compatible server. VS Code offers a model in agent mode only if it supports tool calling. Reference: AI language models in VS Code |
| OpenCode | configure a provider that serves a model on your machine, following OpenCode's own documentation, then add the Studium server: Connect OpenCode |
Building a lesson is the demanding part: every quotation must match its source passage character for
character, and Studium refuses a draft that does not. Playing a lesson asks less: the assistant shows the
text Studium returns and passes your words on. Lessons and progress live in ~/.studium, whatever the
assistant, so a lesson built with one assistant can be played with another on the same machine.
The judge model
Without a judge, nothing is lost: on a written answer Studium shows the reference, you compare and rate your recall, and the verdict says that nobody assessed it. With a judge, the audit steps are assessed by the model, and the verdict says so.
Set the judge in the environment your assistant gives the Studium server, where the commands of this
documentation leave LITELLM_BASE_URL= empty:
| Variable | Value |
|---|---|
LITELLM_BASE_URL |
the address of an OpenAI-compatible server, without /v1: Studium posts to <address>/v1/chat/completions |
LITELLM_API_KEY |
the bearer key, when the server asks for one |
LITELLM_MODEL |
the model name of the audit judge; studium/audit-judge by default |
Studium sends two kinds of request to that address:
| Step | Model name sent | If the endpoint fails |
|---|---|---|
| Audit of an AI's reading or summary | LITELLM_MODEL |
the answer is recorded as not assessed |
| Discussion | studium/runtime-tutor, which cannot be changed |
the step goes on without the model |
The other written answers (a critique, an explanation in your own words, a comparison) are not sent to
this address. They are assessed only by the scene-synth service of a full deployment, which Studium
looks for at http://localhost:8002; without it you compare with the reference and rate your recall.
The other settings are in Configuration.
A non-empty LITELLM_BASE_URL may be a paid route: check what the endpoint charges before you set it.
A judge on your machine
A LiteLLM proxy can give the two names to a local model. An example to adapt, for a model served by Ollama:
model_list:
- model_name: studium/audit-judge
litellm_params:
model: ollama/<model>
api_base: http://localhost:11434
- model_name: studium/runtime-tutor
litellm_params:
model: ollama/<model>
api_base: http://localhost:11434
Start the proxy with this file, then give its address to Studium, for example
LITELLM_BASE_URL=http://localhost:4000.
What you get
get_server_inforeportsmodel_backend:none, orconfiguredwhenLITELLM_BASE_URLis set. Ask your assistant to call it.- Each verdict says who assessed the answer (
assessed_by): the author'skey, ajudge, ornone. The notice on the first screen of a lesson an assistant wrote says whether a judge assesses written answers. - If the judge does not answer, or answers with an error, the lesson goes on and an audit answer is recorded as not assessed.
Limits
- No model that runs on a machine has been tried with Studium, neither as the assistant's model nor as the
judge. What a run checked is the audit request: one request to
<address>/v1/chat/completionswith the model name, the key and a JSON reply format, and what Studium does when it fails. The discussion request was read in the code, not run. - The LiteLLM file above has not been run with Studium.
- With a judge, the server refuses to open a session when it has no secret of its own for learner
pseudonyms. That is the case with
STUDIUM_EPHEMERAL=1, and with a data folder that already held learners before Studium gave each installation its own secret. A data folder created since then needs nothing more. Read Configuration before you setSTUDIUM_PSEUDO_SECRET. - An HTTP 200 from an endpoint does not prove that a model answered.
- Plans, models and limits of an assistant are set by its provider and change over time.