ahwurm/localharness is an open-source coding agent harness for local LLMs. The latest release is v0.23.0. The provided record does not list the items changed in this version. According to the repository description, it is designed so that users can start their own harness or replace closed-source tools. It is written in Python and released under the MIT license.
This tool is used as a coding and research agent by connecting to a local model server or running it directly. It is described as supporting vLLM, llama.cpp, Ollama, and LM Studio, and it lets you define agents in YAML and configure tools, MCP servers, and permission-check procedures. Memory, image generation, browser, and web search are provided as plugins, and the automatic research feature is introduced as still under development. Event logging and records, logs, and performance data are provided as well, and the README states that records and memory are kept locally.
Performance data has to be read within a narrow scope. The README notes that measured baselines differ from runtime to runtime, and that the recorded benchmark run is limited to vLLM. Therefore, performance differences across the other supported runtimes cannot be compared from this record alone. It is useful for looking at how agents for local models are configured and extended, but as evidence for judging performance comparisons it is limited.