Skip to content

A real course from the model you already pay for.

Lerience points the Claude Code or Codex you already have at a folder on your disk and teaches you from it, one module at a time.

This is the real interface, running in your browser on a sample course. The tutor here is a short script; in the app it is your own Claude Code or Codex.

Why this exists

Learning from a frontier model works. Ask one to explain something, argue with it, have it read your attempt and tell you exactly which line is wrong, and you get something close to a patient expert with all evening free.

Then you close the tab. Next week the model has no idea which part tripped you up. Nothing you did was checked, nothing got scheduled for review, and at no point did it have a reason to refuse to hand you the answer. So it handed you the answer, and you learned less than you felt like you did.

Developers already know the shape of the fix, because it is how they work with agents now. Keep the state in files and Git so it outlives the session. Write the rules down in an AGENTS.md. Give the agent tests to run against your work and skills to reach for. It works far better than chat. It is also a rig you assemble yourself, once per subject, out of parts built for shipping code rather than learning it, and even assembled it is a terminal, an editor, and a review schedule you keep by hand.

Lerience is that rig, built once and put behind an ordinary app. The model keeps every capability it has. The rules live in files you are free to read. None of the setup is yours to do, whether you have done it a hundred times before or would never have tried.

What a chat window and a fixed course cannot do

It re-plans around you after every module

The spine holds once you agree to it. The content does not. Only the current module exists in full, and the tutor writes the next one when you arrive, shaped by how the last one went. Pass the checks first try with no hints and the scaffold gaps widen. Need the third hint and a stepping-stone task appears first. Every session opens with the two or three recall questions that are actually due.

journal.md, after a session

Compared two route segments that pointed northeast at different scales. The useful correction was to normalize only when the question is about direction. Left open: decide how the implementation reports a zero vector.

The tutor will not do the work for you

The first rule in the protocol is that it never writes solution code. Learning happens in the gap between the scaffold and the passing checks, so the tutor explains, asks what you think is happening, points at the line, and releases one hint at a time. It will not fill the gap when you ask it to directly. Phase gates open only when you pass them, and every attempt lands in the record.

The tutor, asked for the answer

I won't write it for you. Look at line 14. You divide by the product of the two lengths. What is that product when either vector is [0, 0]?

It builds a picture when prose is not enough

When a module turns on something spatial or dynamic, the tutor writes a self-contained visualization for it, derived from the lesson it has just written, and embeds it in the chapter where the picture belongs. If a misconception shows up mid-session, it can retarget the visual at it on the spot. Lerience renders these in a sandbox with no network and no file access.

The Visual tab in the window above

Vectors and similarity. Drag two arrows. Watch dot product, length, cosine, and distance move together.

Not only code

A course is whatever you ask for. The interview at the start settles the subject, the pace, and how you want to be checked, and the tutor builds around that. Three courses, chosen because they have almost nothing in common:

Programming
The vector course at the top of this page. Every module carries a scaffold with the load-bearing parts cut out and checks you run yourself. The tutor reads your code and points at the line; the checks decide whether it passes.
History
A course on the golden age of piracy in American waters. Nothing to run, so the learner chose at the interview how to be checked: comprehension questions in every module, and a short essay at each phase gate, which the tutor reads and grades against the modules it covers. Recall works the same as anywhere else.
Interview prep
A job description went in as the first input, and notes from each real interview round went in as the course ran, so the modules kept reshaping around what that employer valued. Early modules brushed up the skills with clear deadlines; later ones were practice interviews, with the tutor playing the interviewer.

Runnable checks are the strictest grader Lerience has, and the protocol tells the tutor to say so when a subject cannot produce them. Where it cannot, the tutor grades your written work itself, and you should expect a careful reader rather than a test runner.

Your course is a folder you own

Files and Git history on your disk. App updates never rewrite course content or your record. Pick a file to see what it is for.

CLAUDE.md

The tutor protocol: about 200 lines of plain markdown you can read in ten minutes. It covers the onboarding interview, the open and close of every session, how modules are generated, the hint policy, and the rule the tutor may never break.

The first rule in the file

Never write solution code. The learning happens in the gap between the scaffold and the passing checks. You may explain concepts, ask Socratic questions, review the learner's code and point at the line where the problem is, reveal sealed hints one level at a time.

Before you install

You bring the subscription.
Lerience is free and MIT licensed. The tutor is your own Claude Code or Codex, and it has to be installed and signed in already. Lerience does not bundle, download, or copy provider clients or credentials; sign-in, billing, and terms stay between you and the provider.
Windows and macOS, unsigned, early.
This is a first community preview. Windows may show an unknown-publisher warning. macOS blocks the first launch until you choose Open Anyway under System Settings, Privacy & Security. Lerience verifies its own updates with a signed manifest, an exact size, and a SHA-256 match, and you still approve every download and install.
Runnable checks are the strictest grading it has.
Programming and its neighbours get them. Where a subject cannot produce them, the tutor grades your written answers and essays itself and tells you so, rather than quietly pretending; expect a reader's judgement there, not a test runner's.
The ceiling is the model's ceiling.
The protocol raises the floor and the scripts catch a class of mistakes, but a generated module can still be wrong and the tutor can still misjudge you. The QA step exists because both have happened.

Lerience 0.0.7 for desktop

Windows 10 or 11, x64, or macOS on Apple Silicon or Intel. On a Mac, About This Mac tells you whether it has an Apple chip or an Intel processor. You also need Claude Code or Codex installed and signed in.

Windows
Lerience-Setup-0.0.7-x64.exe140 MB
Mac · Apple
Lerience-0.0.7-arm64.dmg188 MB
Mac · Intel
Lerience-0.0.7-x64.dmg193 MB
Verify
Release notes and checksums
Source
github.com/sqmch/lerience