Pyrefly is a fast type checker and language server for Python developed by Facebook, written in Rust and designed to provide both command-line type checking and IDE integration. The project achieves exceptional performance, checking over 1.85 million lines of code per second and type checking projects like PyTorch 15 times faster than Mypy and Pyright. In IDE environments, rechecks typically complete in under 10 milliseconds after saving a file. Pyrefly serves as the default type checker for Instagram's 20-million-line Python codebase at Meta and has been adopted by major open source projects including PyTorch and JAX, demonstrating its production-readiness at scale.
The tool functions as a comprehensive language server implementing the Language Server Protocol, offering code navigation, semantic highlighting, code completion, hover information, and inlay hints with consistent results across both the command-line interface and various editors. Pyrefly is available as a VSCode extension, Neovim plugin, Zed extension, and supports additional editors through its language server implementation. The project reached stable status with version 1.0.0 and follows a monthly release cycle for minor versions with patch releases as needed for critical fixes.
A distinctive feature of Pyrefly is its built-in support for real-world Python frameworks and tools. It includes native support for Pydantic with model validation and field type inference, Django integration, and pytest fixture navigation and autocomplete, allowing developers to work with these popular tools without additional configuration. The type checker understands the patterns and conventions used by these frameworks out of the box.
Pyrefly addresses adoption challenges through migration tools designed for teams transitioning from other type checkers. The pyrefly init command facilitates migration from Mypy or Pyright, the pyrefly suppress command silences existing errors to allow gradual adoption, and the pyrefly infer command generates type annotations automatically. This approach enables teams to start type checking a single file and expand coverage at their own pace rather than requiring wholesale codebase changes.
The repository demonstrates active maintenance and community engagement. GitGenius tracking data shows a median issue and pull request response latency of 1.3 hours with a mean of 17.0 hours across 2518 tracked items, indicating responsive project management. The most active contributors tracked include yangdanny97 with 4170 events, rchen152 with 1336 events, and kinto0 with 907 events. The project's issue tracking reveals typechecking as the most active label with 1168 items, followed by language-server with 315 items and needs-discussion with 176 items. The repository shares contributors with python/cpython, golang/go, and astral-sh/ruff, indicating cross-project collaboration within the Python and developer tools ecosystem.
