Repolode
lean-dojo

lean-dojo/LeanCopilot

LLM-powered proof automation for the Lean theorem prover, offering tactic suggestions, proof search and premise selection.

Stars
1,341
Forks
129
Open issues
0
Last commit
26 days ago

What is it?

Lean Copilot lets large language models run directly inside the Lean 4 theorem prover to automate parts of writing formal proofs. It provides tactics such as suggest_tactics, which proposes next proof steps, search_proof, which combines LLM-generated tactics with aesop to find multi-step proofs, and select_premises, which retrieves potentially useful lemmas from Lean and mathlib4.

The project ships with built-in models from LeanDojo, downloaded from Hugging Face, and runs them locally through CTranslate2, with or without a GPU. Users can also plug in their own models, either local or remote, through a generic generator and encoder interface or an external API server.

Lean Copilot is a Lean package added via lakefile configuration. It requires a Lean version of at least v4.3.0-rc2 and supports Linux, macOS, Windows and WSL, with prebuilt releases for most platforms.

Key features

  • 01Suggests tactics for the current proof state, optionally constrained by a prefix.
  • 02Searches multi-tactic proofs by combining LLM tactics with aesop.
  • 03Retrieves relevant premises from Lean and mathlib4 with a LeanDojo retriever.
  • 04Runs built-in or custom models locally via CTranslate2 or remotely via an external API.
  • 05Exposes a general LLM inference interface usable beyond theorem proving.

Use cases

  • Getting suggestions for the next tactic while writing a Lean proof.
  • Automatically searching for short multi-step proofs of simple goals.
  • Finding relevant lemmas in mathlib4 when you don't know their names.
  • Building custom LLM-based proof automation or other applications inside Lean.

Quick start

  1. 1

    Add the LeanCopilot require line to your lakefile

  2. 2

    Fetch the LeanCopilot package

    lake update LeanCopilot
  3. 3

    Download the built-in models from Hugging Face

    lake exe LeanCopilot/download
  4. 4

    Build your project

    lake build

Things to know

  • Requires a Lean version of at least lean4:v4.3.0-rc2.
  • On Linux, a lean_exe target needs extra linker configuration to build.
  • search_proof can sometimes produce erroneous proofs.
  • select_premises only covers a fixed snapshot of Lean and mathlib4.

In collections

Topics

FAQ

What is LeanCopilot?
LLM-powered proof automation for the Lean theorem prover, offering tactic suggestions, proof search and premise selection. Lean Copilot lets large language models run directly inside the Lean 4 theorem prover to automate parts of writing formal proofs. It provides tactics such as suggest_tactics, which proposes next proof steps, search_proof, which combines LLM-generated tactics with aesop to find multi-step proofs, and select_premises, which retrieves potentially useful lemmas from Lean and mathlib4.
Is LeanCopilot free to use? What license does it have?
LeanCopilot is open source under the MIT license. Check the license terms before using it in commercial products.
Is LeanCopilot still maintained?
Actively maintained. The last commit to LeanCopilot was on Oct 13, 2026.
What can I use LeanCopilot for?
Getting suggestions for the next tactic while writing a Lean proof. Automatically searching for short multi-step proofs of simple goals. Finding relevant lemmas in mathlib4 when you don't know their names. Building custom LLM-based proof automation or other applications inside Lean.
How do I install LeanCopilot?
Add the LeanCopilot require line to your lakefile Fetch the LeanCopilot package: lake update LeanCopilot Download the built-in models from Hugging Face: lake exe LeanCopilot/download Build your project: lake build
Where can I get LeanCopilot?
The source code and installation instructions are on GitHub: https://github.com/lean-dojo/LeanCopilot

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects