Tutor: an AI coding assistant that runs 100% locally
Tutor is a terminal coding assistant that runs on a local 7 to 9 billion parameter language model, on an 8 GB GPU. It does directly what you already master, teaches what you are learning and measures, without AI, whether learning happened. Cloud models are optional and stay locked until the user allows them.
Client and context
Coding assistants solve things fast, but people who use them all the time may stop learning. Tutor separates what a person wants to outsource from what they want to learn, and behaves differently in each case.
The problem
“I want an AI that does the repetitive work for me but makes me think about what I am still learning — without sending my code off my machine.”
Constraints
- A small model on an 8 GB GPU: it needs guardrails so it does not fail confidently.
- The user's code does not leave the machine by default.
- Learning measurement cannot depend on the model itself.
Decisions and trade-offs
Local model via Ollama, cloud only with permission
Why: Privacy and zero per-use cost. Claude and Gemini join by role (judge, escalation) and only after the user allows it.
Modes chosen by the user's goals
Why: Topics marked “outsource” are executed directly; those marked “learn” switch to guided mode, with hint ladders that never hand over the solution.
Measurement without AI
Why: Spaced reviews (FSRS) where the right answer comes from running the code, and micro-tasks graded by hidden tests.
Guardrails for the agent
Why: Each step is a reversible patch; touching files outside the plan or changing a test to make it pass requires permission.
Research metadata from databases, never from the model
Why: Paper titles, authors and references come from OpenAlex; the model only comments on the abstract.
What we built
- Terminal chat with execute, guided, AI-free practice and review modes.
- Pair mode: the task is split into steps between the AI and the person.
- Relevance-ranked repository map (PageRank over the import graph) and a cached index.
- Academic search on OpenAlex, with ABNT and BibTeX export.
- Mastery dashboard per topic, with no points, streaks or leaderboards.
Results
- 8 GBof GPU memory is enough to run the local model (tested with Ollama)Source: Project README
- 135automated testsSource: Tutor repository, Oct 13, 2026
This case's tile
Non-blank Python lines in each Tutor module.
| Tutor lines of code by module | lines |
|---|---|
| Pedagogy | 1,473 |
| Research | 1,060 |
| Workspace | 1,013 |
| Learning | 630 |
| Models (LLM) | 449 |
| Evals | 237 |
| Adaptive profile | 211 |
Source: Tutor repository · 2026-10-13
Technologies
Want something at this level?
Send the problem, not a spec.