Our take
- Activity
- 19
- Community
- 56
- Issues
- 21
- Pull requests
- 20
+13 stars in the last 30 days
Project history
Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.
Stars
4,542Forks
319Change since Nov 16, 2026: 0 stars·0 forks
Overview
OpenAI Harmony provides shared rendering and parsing for Harmony-formatted conversations, the response format used by OpenAI's gpt-oss models. It helps inference developers construct model inputs and interpret outputs without implementing the format themselves, including its conversation structure and tool-call conventions.
The Rust implementation is also exposed through a Python package. This is most relevant when you operate or build an inference stack that exposes Harmony tokens directly; API and provider integrations may handle the formatting for you.
Key Features
- Render conversations into Harmony-formatted model inputs and parse completion tokens back into messages.
- Represent roles, messages, and conversations with Rust data structures and a corresponding Python API.
- Support Harmony channels, tool-call preambles, tool namespaces, and structured outputs.
- Use built-in encodings, including the gpt-oss Harmony encoding.
- Run the core implementation in Rust, with Python bindings provided through PyO3.
- Install the Python package from PyPI or use the Rust crate from the repository.
Use Cases
- Inference-platform developers can format prompts and decode completions when serving gpt-oss directly.
- Rust application developers can integrate Harmony rendering and parsing into model-serving or orchestration code.
- Python developers can build custom inference workflows with the package's message and conversation helpers.
- Teams testing format compatibility can use the Rust and Python test suites to check their integrations.
Getting Started
Install the Python package:
pip install openai-harmony
For Rust setup, usage details, and development instructions, see the repository README.
Considerations
- Harmony formatting is important for using gpt-oss correctly in a custom inference solution. If an API or hosting provider handles the format, you may not need this library.
- The Rust crate is the core implementation. Building the Python package from source requires a Rust toolchain and maturin.
- The README recommends Python 3.8 or later for local development. Check the project documentation for further language-specific guidance.
Found this useful?
Share it with someone who would like harmony.