Agents work in parallel
A planner splits the theorem into lemmas that Lean accepts as a valid plan, and several provers work on them at once. You get a verified proof, or a report on what is still open.
AI writes the proof. Lean checks it.
State the problem in VS Code. Agents work until Lean says yes.
Join the demo list
Product preview coming soon
A screenshot of MathIDE in VS Code will appear here.
A confident model is not a proof.
AI can now help with real mathematics, but it can sound sure about a wrong argument, and checking by hand eats the time you hoped to save. Lean already checks every step. MathIDE keeps it in the loop, so only checked results are called proved.
-
Verified means a clean build.
A result is labelled verified only when Lean builds it with no
sorryand only standard axioms. Anything else is marked as an unverified suggestion. -
Your statement is the contract.
You state the problem once. Agents cannot change it, and every result is shown next to it, because Lean can prove the wrong thing.
-
AI use is logged.
Every agent change is a git commit, and the log records what was sent, so you can disclose tools and prompts in your paper.
-
It runs without you.
Agents work in a separate branch under a fixed budget, with access only to your Lean project. You read the result when it suits you.
Interactive example
See what your proof rests on.
Click a theorem to read what it says in plain language. This is an illustrative map of the classic proof that the square root of 2 is irrational.
In plain language
Example only. MathIDE would build this kind of map from your own Lean project.
From proof to paper in one place.
A paper that matches the proof
Agents draft the LaTeX from statements Lean has already verified, in the style your field expects.
\begin{lemma} For every natural number $n$, we have $2n = n + n$. \end{lemma} % checked in Lean: % two_mul_eq_add
No new workspace
It runs in the VS Code and Lean project you already use. No export step, no second tool to learn.
Questions mathematicians ask.
Can I use it today?
No. MathIDE is still being designed. Joining the list tells us what to build first, and you will hear from us when there is something to try.
Do I have to watch it work?
No. You set the problem and a budget once. Agents work in a separate branch and report back when Lean accepts a proof or the budget runs out. A short list of doubts about your statement waits in your inbox.
Does the AI become an author?
No. AI is an instrument here. You stay responsible for the statements and the paper, and the log of AI contributions supports disclosure.
What do I need to use it?
VS Code and a working Lean 4 setup. The plan is to use your own AI provider key, so you choose the provider and see what is sent.
What happens to my unpublished work?
You choose once per project which files may be sent to your AI provider, and everything sent is logged. Text sent to a provider goes to that provider, so read its retention terms before use.
Why not use a general coding agent with Lean?
You can, and it works if you are comfortable in a terminal. MathIDE puts that workflow in the editor and ties the proof map and the paper to the same verified results.
Join the demo list.
Tell us how you write Lean and LaTeX today. Early replies decide what we build first, and a person reads every one.
We will email you about the MathIDE demo and nothing else. Details are on the privacy page.