A working miniature of a practice session, running the app's rules in your browser. Get it wrong on purpose, then retry, and see what does and does not count.
Practice, Calculus for machine learning
ModelLocal model
Calculus for machine learning, Chain rule
Demonstration item, no hints
Differentiate f(x) = (3x2 + 1)4
Hints off for this item
Pick an answer and submit it. Your attempt is saved before it is checked, and only the result of the check can move the state.
It decides what comes next.
Open the app and it tells you the most useful thing to do with the time you have, across every subject, and why.
Tonight, 45 minutes
Ranked by rule, not by a model
ReviewHTTP basics
10 min
Overdue by 2 days
PractiseChain rule
20 min
Unblocks 3 concepts
LearnPreterite: regular verbs
15 min
Spanish is behind its weekly hours
DemonstrateMaking requests
20 min
Checkpoint in 3 days
Confirm2 provisional markings
5 min
Short, batched
The last two did not fit tonight and move to tomorrow. The same state and the same clock always produce the same plan, and every suggestion says why it is there.
Reviews that space out
Pass a review and the next one moves further away. Miss one and it comes back in 2 days.
7d16d35d75d
Reviews fall 7, 16, 35 and 75 days after a concept is demonstrated.
Blockers, named
The planner never schedules a blocked concept. It schedules whatever is blocking it.
Chain rulescheduled tonight
Backpropagation
Training a network
One system. Every subject taught its own way.
Mathematics is not learned like a language. The record and the five states stay the same; the practice, the checking and the reviews change with the subject.
“Build and deploy a useful application in eight weeks.”
First discipline to ship
A typical 45-minute session
5 minReview a due concept
10 minIntroduction by questions
20 minImplement with tests
5 minExplain your solution
5 minSummary
Orange steps produce evidence. The rest is exposure.
You practise with
Exercises, implementations, debugging, projects
Checked by
Unit tests and output comparison, run on your machine
Demonstration looks like
Implement from a specification with every test passing
Review looks like
A small new implementation, or a bug to find
Your record outlives your model.
The app owns the learning record. Models and agents are interchangeable workers on it, so you can switch whenever a better one appears and lose nothing.
Your learning recordOne SQLite file, on your computerThe model you bring
Anthropic
OpenAI
Google Gemini
OpenRouter
OllamaThe agents you use
Claude Code
Codex
Any MCP clientPlain Markdown, any time
Answering now: AnthropicSwitch the model and the record does not change. Agents reach it through the MCP server.
A model's output is a proposal
Everything a model returns is checked against a schema before the app decides what to write. A failed call cannot damage your record.
You see where it goes
The active model is always on screen, and what each request cost is counted, with caps you set.
API keys are billed separately
A chat subscription is not an API key. You pay your provider by usage, or nothing at all with a local model.
Fifteen Kircher skills: plain SKILL.md files that every agent in the app follows, from mapping a field to making it stick. Open source is planned; for now the repository is private while we bring the skills up to the bar we set for them.
---
name: assessor
description: Writes fair questions and quizzes that test
whether someone can actually do a thing, and marks
answers honestly against an explicit rubric with a
stated confidence, naming the specific misconception
behind a miss.
---
# Assessor
You find out whether someone can do a thing, and you
tell them the truth about it in a way they can act on.
Two halves: writing questions that are fair, and
marking answers that are honest.
## Writing a quiz
### 1. Set it up first
Never write a quiz before you know what it covers,
how many questions, how hard, and which kinds.
Free while it is in beta.
Everyone in the early beta gets the whole app, unlocked. After the beta: a free version with one track, and one payment for everything. You bring the model, so we never charge for tokens.
Early beta
Now
Freewhile the beta runs
The whole app, every capability unlocked, for everyone who joins the beta.
Unlimited tracks
All nine agents and the Kircher skills
Lessons, practice, checks, projects and reviews
The five states and the full evidence log
Today, the planner and the schedule
What you tell us decides what gets built next.
Free
₹0after the beta
The complete learning loop, on one track.
One track
Lessons, practice, checks and reviews
The five states and the full evidence log
Today and the planner
Notes and your own sources
Full version
One paymentprice to be announced
Pay once and keep every capability. No subscription, no renewal.
Unlimited tracks
Everything in Free
Every agent, skill, tool and connection the app has