everyharness

everyharness is an offline-first CLI and TUI tool for managing local machine learning models. It provides a model registry that can handle local files, Hugging Face, Ollama, and Python callables. Users can register a model and run it through a harness, with a consistent flow of add, run, serve, and train. The tool is designed for people who keep models on disk and want a single interface instead of multiple tools.

It supports various model types including tabular (sklearn/joblib), embeddings, LLM (Ollama, optional GGUF/HF), vision, diffusion, speech, and computer. The maturity of these features varies, with tabular being solid and others being usable or thin. The tool includes a Textual TUI and can generate agent prompt packs for coding agents. It is plugin-shaped, allowing for community plugins via everyharness-* packages.

everyharness is distributed on PyPI and is available for macOS and Linux (v1). It is licensed under Apache 2.0. The tool emphasizes offline operation, with an environment variable EVERYHARNESS_OFFLINE=1 that hard-blocks outbound network access. It also includes a --trust-pickle flag for loading pickles, with a recommendation to prefer joblib from trusted sources.