The Problem
Building and deploying AI models typically requires stitching together separate toolchains: a language runtime for model logic, a serving layer for inference, and a kernel library for performance. The Modular Platform addresses this by unifying model development and deployment under one umbrella, with the MAX framework for serving and the Mojo language for low-level performance work. This repo is the open-source home of those components.
What This Does
This is a portfolio of nine self-contained projects, not a single codebase. The three substantial ones: max/ (5,844 files) — the MAX framework: accelerator kernels (max/kernels), an OpenAI-compatible inference server (max/python/max/serve), and model pipelines (max/python/max/pipelines). KGEN/ (2,329 files) — the Mojo compiler and its tooling, including a Jupyter executor (KGEN/tools/mojo-jupyter-executor/main.cpp) and a REPL entry point. mojo/ (1,344 files) — the Mojo standard library (mojo/stdlib) and examples.
The remaining projects are supporting infrastructure: AsyncRT/ is a C++ async runtime with a work queue and allocator; bazel/ holds build configuration; Support/ and docs/ contain utilities and documentation. Recurring techniques across projects include Bazel as the build system, C++ for runtime internals, Python for the higher-level API surface, and GitHub Actions for CI.
How To Use It
The README directs users to external quickstart guides for both MAX and Mojo. The repo's own README.md and CONTRIBUTING.md are the authoritative entry points; no local build instructions are in the root.
Setup — no root-level install command exists. The presence of pyproject.toml files (e.g., KGEN/tools/mblack/pyproject.toml) and package.json files (e.g., KGEN/test/mojo-lsp-server-node/package.json) implies pip/uv for Python components and npm/yarn for Node components, but there is no unified bootstrap script.
Configuration — the repo uses .bazelrc and .bazelversion for Bazel builds, and .pre-commit-config.yaml for linting. No environment variables are documented in the root.
Running it — for the Mojo compiler, the entry point is KGEN/tools/mojo-repl-entry-point/main.cpp. For the Jupyter integration, KGEN/tools/mojo-jupyter-executor/main.cpp. For the MAX inference server, the README points to max/python/max/serve.
Clone the repo
git clone https://github.com/moses-y/modular cd modular
See the README for external quickstart links
No local build commands are documented in the repo root
Real-World Use
A typical scenario: a team serves a Hugging Face model with the MAX inference server (max/python/max/serve), which exposes an OpenAI-compatible endpoint. For custom kernels or performance-critical paths, they write Mojo code using the standard library in mojo/stdlib and compile it with the toolchain in KGEN/. The AsyncRT/ runtime handles the underlying async execution for concurrent inference requests.
Code Health & Issues
Static analysis (tree-sitter parsing) found 380 functions and 118 classes across the repo, with 4,817 test files and 1,300 documentation files. No structural red flags: tests, CI, license, and lockfiles are present; no committed secrets detected. The analysis is deterministic and based on file structure, not opinion. Med — dependency drift: @types/node (^24.0.10), typescript (5.8.3), and vscode-languageserver (^9.0.1) are all behind current major versions, per the package.json files. Low — the repo is a monorepo with nine distinct projects; there is no unified build or test entry point, which complicates local development.
The Bottom Line
This is the official open-source repository for the Modular Platform, containing the MAX framework and Mojo language. It is substantial and well-maintained, with strong test and CI coverage. It is suited to teams building AI infrastructure who want the Mojo/MAX stack, but the lack of a unified local build path means most users should follow the external quickstart guides rather than building from source.