PenguinHarness
Open-source AI agent builder for macOS, Windows, Linux and self-hosted servers, with local or cloud models and automated evaluation and tuning.
PenguinHarness runs on your computer or server and uses AI agents to build runnable agent applications from plain-language requirements. It's for people developing AI apps who want the same tool to handle creation, evaluation, tuning and deployment. The project is open source under Apache 2.0.
Its self-improvement loop tests an agent against benchmarks, examines scores and execution traces, then revises its prompts and Skills. Parallel evaluators can assess the same agent, and snapshots let you restore an earlier state if a revision performs worse. Changes stay within the agent's workspace and Skills; the loop doesn't rewrite the harness core.
You can use a desktop app on macOS, Windows or Linux, or run the server yourself with Docker and access its web interface. A CLI and SDK expose the same engine for scripted work. It supports local and online models through OpenAI-compatible endpoints, with provider connections including DeepSeek, OpenAI, Anthropic and OpenRouter. You can switch models without rewriting the agent. The app and its stored data run locally; choosing a cloud model sends model requests to that provider.
The interface includes agent management, multi-session chat, scheduled tasks and isolated subagents. Built-in Skills cover browser automation, data analysis and software development. Trace views record model requests and tool calls, including failures, timing and token use. Tool calls pass through approval checks, and multi-user projects keep each user's data separate.
