Jichen Feng

Upenn Chem Grad, Princeton Research Assistant.
Developing MathCode: a frontier mathematical coding agent

[Jichen Feng @jcfeng27]

Higher efficiency, lower cost.

[Yifan Zhang @yifanzhang_]

Introducing MathCode 0.2.0: maximize prompt-cache hit rates and reduce API costs by up to 90%.

Project Page: https://t.co/gZZJqxYkd0

[GitHubDaily @GitHub_Daily]

想用 AI 去做数学定理的形式化证明，大多数工具没这个能力，想自己搭一套流程又门槛极高。

无独有偶，发现 MathCode 这个开源项目，用自然语言描述数学问题，它能自动转成 Lean 4 定理并尝试完成形式化证明。

简单说就是一个终端里的数学证明助手，输入一句话比如「证明偶数的平方是偶数」，它就会自动完成从形式化到证明的全过程。

GitHub：https://t.co/OdWgxoqmIm

同时集成了 Lean LSP，能自动从 Mathlib 数学库中检索已有引理来辅助证明，遇到编译错��还会自动修复重试，最多迭代十轮。

还能生成 Obsidian 知识图谱，把定理和引理之间的依赖关系可视化。支持多规划器并行，同时跑多条证明策略找最优方案。

如果你对数学形式化验证感兴趣，或者正在用 Lean 4 做定理证明觉得手动流程太繁琐，这个工具值得试试。

[Jichen Feng @jcfeng27]

Will stay in SF for several days. "Milestones of Autonomous Mathematics" workshop hosted by ICARM and Principia labs will extremely interesting.

[Jichen Feng @jcfeng27]

Welcome MathCode users to join our discord to share your feedbacks and using experience! https://t.co/uuLpD1PlVQ

[Yifan Zhang @yifanzhang_]

Introducing MathCode 0.1.0, Stronger than ever.

It now supports customized tools and skills. PRs are welcome!

https://t.co/gZZJqxXMns

[Shihan Qin @Shihan_Qin]

Introducing MathCode v0.1.0 — our largest update yet.

Supporting Tree of Subgoals, TheoremLib, AxiomLib, Guided Planning, Blueprint and Tools & Skills！

Try it at: https://t.co/J192GmRCm0

[Jichen Feng @jcfeng27]

@enjoyingthewind @yifanzhang_ @Shihan_Qin @BaoqingYue1 @Young_AGI Now we have it!

[Jichen Feng @jcfeng27]

MathCode v0.0.3 — Lean proofs as an Obsidian knowledge graph. The agent reads the vault, searches Mathlib via LSP, writes proofs, compiles, retries autonomously. Multi-planner runs parallel strategies. Every lemma feeds back into the graph.

https://t.co/AdW0a9DnEW

[Jichen Feng @jcfeng27]

Releasing MathCode v0.0.2, a frontier AI agent that formalizes and proves math in Lean 4.
Now defaults to Codex, supports Claude API and MiniMax. Mix backend: Codex for chat +Claude for math also available.

Natural language → Lean 4 theorem → formal proof.

Try at: https://t.co/Wcyux6R5zE

[Jichen Feng @jcfeng27]

Thank you so much for supporting MathCode!

[Xiuyu Li @sheriyuo]

Excited to see these new features. Great work🥳

[Jichen Feng @jcfeng27]

7/n Tools and Skills: MathCode now supports tools and skills. We are welcoming MathCoders to share their wonderful skills and tools and make contributions to the MathCode community! We appreciate everyone’s supports for MathCode!

[Jichen Feng @jcfeng27]

1/n Releasing MathCode v0.1.0.
Our biggest update, adding:
- Tree of Subgoals
- TheoremLib
- AxiomLib
- Guide of Plans
- Blueprint
- Tools and Skills.

https://t.co/jbYJgpE7dC

[Jichen Feng @jcfeng27]

MathCode v0.0.3 — Lean proofs as an Obsidian knowledge graph. The agent reads the vault, searches Mathlib via LSP, writes proofs, compiles, retries autonomously. Multi-planner runs parallel strategies. Every lemma feeds back into the graph.

https://t.co/AdW0a9DnEW

[Jichen Feng @jcfeng27]

6/n Blueprint: MathCode now supports writing the markdown files which could be rendered to blueprints showing the proving scheme of each theorem to lemma.

[Jichen Feng @jcfeng27]

5/n Guide of Plans: MathCode now can learn from the conversations with you, and develop the prove plans from it. Will be very useful for complex math problems that needs the guides and insights from smart humans.

[Jichen Feng @jcfeng27]

4/n AxiomLib: the wildest feature we added this version . Chat with your MathCode, and it will create your personal axioms that you mentioned. It could be anything: math, physics or even narrations. MathCode helps you create your own thoughts world and proves the logic inside.

[Jichen Feng @jcfeng27]

3/n TheoremLib: MathCode now can store your proved theorems and add them to your personal TheoremLibs. These theorems will be recognized and could be import in the future proofs. This makes a step-by-step big theorem proving possible.

[Jichen Feng @jcfeng27]

2/n Tree of Subgoals: MathCode now supports solving math problems by adding subgoals and distributing to multiple agents, which will significantly increase the power and speed of MathCode. Multiple layers are also available.

[Jichen Feng @jcfeng27]

@enjoyingthewind @yifanzhang_ @Shihan_Qin @BaoqingYue1 @Young_AGI Thank you for supporting MathCode!

[Jichen Feng @jcfeng27]

@enjoyingthewind @yifanzhang_ @Shihan_Qin @BaoqingYue1 @Young_AGI Definitely possible, will soon add features in the future release

[Shihan Qin @Shihan_Qin]

Introducing MathCode v0.0.3!

Lean proofs built as an Obsidian knowledge graph. The agent mode can read the vault, queries Mathlib through LSP, writes proofs, compiles, and retries autonomously. A multi-planner explores parallel strategies now!

https://t.co/IA9tNJcurg

[Jichen Feng @jcfeng27]

This is insane!

[Jeanne @prompterminal]

I've combined Manim @NousResearch's Hermes Agent skill + @yifanzhang_'s Math Code.

Math Code executes the proof on a problem called Jordan's Lemma and Hermes Agent with @claudeai Sonnet 3.7 directs Math Code, writes a script, gets Manim to render an explanatory video.

Jordan's Lemma: It helps to simplify high-dimensional problems by breaking them down into manageable, 2D pieces. The lemma is the underlying mathematical reason why Grover's Algorithm (useful in quantum) works by rotating within a 2D plane.

Rendered explanation video below:
https://t.co/9sKcpiz0sd