Lean Runtime

Run Lean proofs from Python or a single .lean file—without creating a throwaway Lake project or rebuilding the same dependencies on every machine.

Lean Runtime discovers the exact Lean environment a project needs and reuses a downloadable copy when one is available. It returns structured Lean results with a record of the toolchain and dependencies that were actually used.

Status: V1 beta. The local backend runs trusted Lean, Lake, and package code; it is an orchestration boundary, not a security sandbox.

Install

python -m pip install lean-runtime

Lean Runtime manages its own Elan installation on macOS and Linux. Windows currently requires LEAN_RUNTIME_ELAN.

Run one Lean file

Inside an existing pinned Lake project, just pass the file:

lean-run MyProject/Main.lean

Standalone files do not need a throwaway Lake project or dependency declaration:

import Mathlib

example : 2 + 2 = 4 := by norm_num

lean-run Main.lean

When no explicit context or pinned Lake project exists, lean-run analyzes imports, ranks a bounded set of exact environments from its bundled catalog, and asks Lean to check each candidate. The successful exact lock is retained by Runtime. Pin it for portable reuse whenever desired:

lean-run Main.lean --lock-out environment.lock.json
lean-run Main.lean --lock environment.lock.json

The initial catalog covers core Lean v4.32.2 plus Mathlib v4.32.2, v4.31.0, and v4.30.0. Runtime first tries its local store and downloadable environment libraries, then builds the exact source environment when necessary. Use --no-source-build to forbid that potentially large fallback or --offline to use retained environments only.

Explicit frontmatter remains available when the desired context is already known:

-- /// lean-runtime
-- requires = ["mathlib@v4.32.2"]
-- ///

import Mathlib

The same context can be supplied from the command line:

lean-run Main.lean --with mathlib@v4.32.2

Create an exact lock from an explicit dependency for CI without changing the file:

lean-run Main.lean --with mathlib@v4.32.2 \
  --lock-out environment.lock.json
lean-run Main.lean --lock environment.lock.json

Python

Configure an environment once, then use it repeatedly:

import lean_runtime as lean

env = lean.setup(["mathlib@v4.32.2"])

result = env.check(
    """
    import Mathlib
    example : 2 + 2 = 4 := by norm_num
    """
)
result.raise_for_error()

Rejected proofs carry parsed diagnostics:

result = env.check(broken_proof)

for error in result.errors:
    print(error.file, error.line, error.message)

result.raise_for_error()  # raises LeanCheckError with the same detail

Core-only work does not need a dependency:

core = lean.setup(toolchain="v4.32.2")
core.check("example : 2 + 2 = 4 := rfl").raise_for_error()

Batch and asyncio APIs reuse that prepared environment:

results = env.check_many(generated_proofs, concurrency=8)
results = await env.check_many_async(generated_proofs, concurrency=20)

Local projects use the same setup pattern while retaining mutable-project semantics:

project = lean.setup(project="./my-project")
result = project.check_file("./my-project/MyProject/Main.lean")

One-shot helpers are available when setup reuse is unnecessary:

result = lean.check(source, deps=["mathlib@v4.32.2"])
result = lean.check_file("./my-project/MyProject/Main.lean")

When you need evidence rather than extra setup, the operations CLI can verify, explain, compare, and measure the same exact contexts:

lean-runtime verify research-stack --offline
lean-runtime compare previous.lock.json environment.lock.json
lean-runtime profile research-stack Main.lean --repeat 5
lean-runtime matrix compatibility.toml Main.lean

Use lean-run Main.lean --explain to inspect context routing without executing Lean, and --timings to expose preparation versus execution time. Successful ordinary checks remain one concise line.

Friendly references remain exact: use mathlib@VERSION, leancert@VERSION, owner/repository@REVISION, or the explicit github:owner/repository@REVISION form. Bare floating package names are never accepted.

Share environments

A project is your ordinary Lake repository. Its environment is the exact Lean version, dependencies, and build configuration needed to use it. A downloadable environment is a ready-to-use copy that collaborators and CI can fetch instead of rebuilding Mathlib.

Environment libraries may be public or private. For example:

lean-runtime --library ghcr.io/owner/lean-environments download environment.lock.json
lean-runtime build-and-publish environment.lock.json \
  --publish-to ghcr.io/owner/lean-environments

For an already-built executable, Lean Runtime can also create a verified ready-to-run program. It opens without rebuilding the project, can be saved as a portable copy, and can be shared through a public or private program library. See Ready-to-run programs.

Technical details

The simple API is backed by exact Git commits and trees, Lake-resolved locks, platform-aware content-addressed environments, atomic cross-process builds, downloadable environment reuse, replayable provenance, verification, and trusted publishers. The libraries use OCI-compatible storage internally, but users do not need Docker or container concepts. Advanced protocol details remain in the architecture documentation.

Documentation

• Getting started
• Python API
• lean-run and operations CLI
• Managed environments
• Local Lake projects
• Portable copies and environment libraries
• Ready-to-run programs
• Architecture
• Trust and limitations
• V1 release case study

License

Apache License 2.0.