ResearchHarness
A lightweight, general-purpose harness for tool-using language model agents, meant as a fair execution substrate for agent benchmarks, a reference baseline, and a personal assistant runtime.
ResearchHarness is designed to be general, stable, fair, lightweight and feature-complete, serving three roles: a fair execution substrate for agent benchmarks such as ResearchClawBench, a reference baseline for comparing future harness designs against, and a lightweight personal-assistant runtime usable directly for coding, file work and report writing. It supports OpenAI-compatible models alongside providers such as Gemini, GLM and Qwen, with a unified interface for tool calling.
It offers both a command-line interface and a Python API, an OpenAI-compatible API deployment mode so it can itself be called like a hosted model, and a local frontend UI. The project's stated goal is not novelty but an inspectable harness that is easy to run, easy to compare against alternatives, and easy to extend.