Distribute and run LLMs with a single file.
C++ cross-platformggufllama-cpplocal-ailocal-inferencelocal-llmopen-source-aisingle-file-executablespeech-to-text
AGENTS.md
This file provides guidance to Claude Code when working with this repository.
Project Overview
Llamafile combines llama.cpp, whisper.cpp, and stable-diffusion.cpp with Cosmopolitan Libc to create single-file executables that run LLMs locally across Windows, macOS, Linux, and BSD without installation.
Quick Reference
# Initial setup (run once after clone)
make setup
# Build (always use cosmocc make, not system make)
# Adapt `nproc` to the OS where you are building, (e.g. `sysctl -n hw.physicalcpu` on mac)
.cosmocc/4.0.2/bin/make -j $(nproc)
# Run tests
.cosmocc/4.0.2/bin/make check
# Clean build outputs
.cosmocc/4.0.2/bin/make clean
# Reset all submodules (warning: removes local changes)
make reset-repo
Key Directories
| Directory | Purpose |
|---|---|
llamafile/ |
Core library (edit directly) |
llama.cpp/ |
LLM inference (submodule, edit directly then convert to patches) |
whisper.cpp/ |
Speech-to-text (submodule, edit directly then convert to patches) |
stable-diffusion.cpp/ |
Image generation (submodule, edit directly then convert to patches) |
*.patches/ |
Patch directories for submodules |
o/ |
Build outputs |
Important Notes
- Always use
.cosmocc/4.0.2/bin/make, not system make - Run
make setupafter cloning or updating submodules - Submodule changes require patch files (see skill for workflow)
Detailed Documentation
For comprehensive build, architecture, development, and testing documentation, ask Claude about "how to build llamafile" or "llamafile development workflow" to load the llamafile skill.