# MathCode cuts Lean compile checks to ~0.4s after warmup > Source: > Published: 2026-08-17 07:41:07.295580+00:00 [Hacker News](https://math-ai-org.github.io/mathcode/) ### MathCode cuts Lean compile checks to ~0.4s after warmup Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated. MathCode drives Lean 4 compilation check latency down from 30 seconds to 0.4 seconds via a persistent language server, making real-time, closed-loop formal math verification viable for LLM agents. For production systems, this ultra-low latency enables agents to instantly self-correct using structured LSP diagnostics, decompose complex proofs into parallelized subgoals, and dynamically cache verified theorems for future reuse. This shifts mathematical reasoning pipelines from slow, asynchronous batch processes into highly responsive, deterministic verification engines. MathCode is a terminal-based AI agent that converts plain-language math problems into Lean 4 formal proofs with a persistent REPL, cutting compile-check latency from ~30s to ~0.4s after warmup. This enables real-time, agentic theorem proving with reusable libraries, parallel subgoal decomposition, and Obsidian knowledge graphs, making formal math verification practical for interactive and production use. ### AI vs. AI Debate “The summary overstates 'instant self-correction' and omits the macOS/Linux dependency, bundled runtime overhead, and the requirement for manual CLI setup.” “The summary appropriately focuses on the core architectural breakthrough of a 98% latency reduction that enables practically instantaneous agentic self-correction, while omitting minor platform-specific dependencies and standard setup configurations that do not alter the system's primary technical contribution.”