Market & Competition

Klar is an agent first learning experience for school, work and life. It gets to know a person, their material and how they learn, then teaches them proactively through generated interfaces. The tools other companies sell as products are features inside that experience.

Most comparisons put every AI learning product in one list. That hides the actual structure. There is a layer people live in, and a layer they visit.

A study tool cannot become Klar without deleting the assumption it was built on. It exists because the general agent exists.

Students are the first customer, not the market. They are the harshest testers: they use it or they do not, they pay or they do not, with no procurement and no company politics in between. That pressure builds the better product.

The same agent applies at work and to anyone learning anything. The engine is the same: know the person, know the material, act proactively so they get capable as fast as possible.

A general assistant is graded on the answer. Faster, more complete, less friction. That is the correct target for a general assistant and they are very good at it.

Klar is graded on the person. Did they retain it. Can they apply it under pressure. Did they need us less this week than last. Reaching that target sometimes requires the opposite behaviour: withholding the answer, asking first, making them attempt it, returning to it days later unprompted.

This is not a tone or a system prompt. It is a different thing to be good at, and it conflicts with what a general model is built to do. A general model that withholds answers is a worse general model.

A chatbot optimises for giving you an answer. Klar optimises for whether you actually learn. Sometimes that means not giving you the answer at all.

The usual objection is what already exists. Coursera and Khan Academy have been teaching online for over a decade. ChatGPT has Study Mode, Claude has learning mode, Google has Guided Learning, study notebooks and NotebookLM. Every large lab has shipped something in this direction, and most of it is free. Below them sits an endless tail of niche study tools launching every week, like Alice.tech, Memmo and others.

Students know these exist and have access to them. They switch to Klar anyway, and pay for it. That is the answer to the objection, and it is an observed result rather than an argument.

The reason is structural. A mode is a feature inside a product built for something else, competing internally for focus, roadmap and brand. Klar is the whole product: the memory of the learner, the proactivity, the generated interfaces, the ecosystem around it and a brand that says one thing clearly.

Finished courses are the same story. Structured courses are a feature inside Klar rather than a product beside it. The system authors them, and the agent teaches from them as context on the person's material. Every course built makes the layer denser, and that density is the data that compounds.

Claude shipping a legal feature does not remove Legora. Ownership of the full experience beats a mode inside someone else's.

There is a reference case for this. Nokia had the better hardware, the wider distribution and the larger market. Apple owned the whole picture: the operating system, the industrial design, the store, the brand, the way it felt to use from the first second. The integrated experience won, and it did not win on specifications.

Learning is the same shape of problem. A person does not evaluate a model benchmark or amount of features. They notice whether the thing understands them, whether it feels considered, whether they want to open it again tomorrow. That is decided by the whole surface: the interface, the memory, the copy, the tone, the brand, the moment it chooses to interrupt them.

Building that requires taste, and taste is not a feature that can be added later. It is why we build product, design and engineering as one function rather than three, and it is the part of this that is hardest to copy from an adjacent market.

The mistake is to treat capability as something to display. Surfacing more tools and more features is the easy signal of progress, and it costs the user clarity. The work in AI is the opposite: the best interface is the one that disappears. Everything the person needs, generated at the moment they need it, and nothing else on the screen. Invisible and more capable, not visible and more crowded.

We do not train frontier models and we do not want to. Klar orchestrates them and picks the right one per task. When the labs ship a better model, Klar gets better the same week, at no research cost.

The consequence is that the fastest moving part of this industry works for us rather than against us. The part we do own, the teaching, is not a thing the labs are optimising.

The learner model.

What a person knows, how they got there, what makes it stick for them. Portable across their studies, their job and whatever they pick up next. It cannot be bought or prompted into existence.

The agent harness.

Teaching is a capability we train and evaluate, not a prompt we write. Every cohort gives us evals on whether an intervention actually worked. The agent gets measurably better at teaching proactively and adaptively over time. Measured against outcomes rather than preferences: did the difficulty land, did the person get there in fewer turns, did it still hold a week later.

The habit.

The layer people open by default is the hardest position to take and the hardest to lose.

Downstream optionality, not the plan: nobody else holds this. The study tools above us lack it, and so do the frontier models below us. It is the one part of the stack that cannot be trained from the open internet, and it could eventually be sold in either direction.

A chat agent is used to automate. Automating is the easier half, and it is the half being built today.

Klar augments. It turns any information into capability. Meet the learning agent.