Penguin Harness

This platform is an open source, local first harness for building AI agents, with a command line interface and a web interface that need no code to use. The vendor describes it as a self improving harness: an agent is handed a requirement in plain language and builds, tests and deploys a working AI application, and the harness can then evaluate and improve the agent over repeated rounds. The site shows an example in which one sentence produced a retrieval augmented generation app that answers Claude Code questions with cited sources, with token cost of about $0.02 on DeepSeek V4 Pro. The self improvement loop uses an Optimizer that runs several Evaluators in parallel on a target agent, scores it against benchmarks, finds where points were lost and upgrades the agent from one version to the next, taking a snapshot before each round. An SDK lets an agent build complete agent applications, and Skills let an agent optimize itself. The vendor emphasizes a deliberately minimal toolset to reduce tool calls and tokens, and publishes its own comparisons with Claude Code and OpenAI Codex on cost and accuracy. It connects to more than 1,000 models through major providers, runs on Linux, Windows and macOS, and is licensed under Apache 2.0. A desktop app, version 0.2.13, and a command line install are offered. The software is free and open source. The vendor also advertises official token packs under the Penguin Go name, with a 50 percent discount on Gemini models, though no price list for them is published.