PenguinHarness is a unified, stable open-source framework for agent self-evolution (recursive self-improvement, RSI). The project address is https://github.com/Prism-Shadow/penguin-harness。其核心理念是：模型保持固定，由 harness evolution, specifically including prompts, Skills, tools, and hooks. It runs on the user's own computer or server and can be used as a desktop application, Web App, CLI, and SDK.
Main capabilities
- Integrate all RSI algorithms into one framework: Default, OPRO, APE, ACE, and AWM are provided as plugins and run on the same set of built-in Benchmarks, scoreboard, and snapshots.
- Build an agent application in one sentence: after describing the requirements, the agent writes the application, pages, and run instructions itself.
- Low operating cost: the toolset is deliberately kept lean and tuned for open-source models; it is said to achieve the best accuracy on a data analysis suite at 1/70 of the cost of Claude Code.
- Agents evolve each other: after selecting an algorithm and Benchmark, one agent improves another round by round; each candidate version is snapshotted before changes, and all scores go into the Evaluation Center.
- Each new project includes seven Benchmarks, run in Docker as Harbor tasks: PenguinHarness Benchmark Sec A to Sec E (from rag-bench-essential, DeepSWE, AutomationBench, Terminal-Bench-Science, and Terminal-Bench) as well as GDPevo Train and Test. More Benchmarks are located in penguin-harness-benchmark.
- Demo example: after a decision-making agent is evaluated, has its working method corrected once, and is retested on the same model, its score increases from 84 to 100, and the cost of each run is displayed on the screen.
One sentence, one application The result example given in the article is a legal review application: the user uploads contracts, review is tracked file by file, and a report containing risk clauses, explanations, and suggested revisions can be downloaded. The model invocation cost for five example reviews is $0.0423.
What the agent can do
- Replies can carry UI: options, forms, steps, Mermaid diagrams, and live widgets (A2UI).
- Each harness runs the same task on the model it is usually paired with.
- Built-in plugins: the agent can also write and optimize its own Skills (see the documentation).
Supported models The documentlists the model families that the project focuses on supporting, including closed-source models and open-source models, and explains that the Models page in the application lists each variant and its provider. The article mentions that OpenCode Go provides Claude Haiku 5.5; Fable 5.1, Opus 5.5, and Sonnet 5.5 can be obtained through Anthropic and OpenRouter; Hunyuan 4 uses hy4-preview. Custom OpenAI protocol endpoints are also available, covering 1000+ online and local models.
How to use
- Download the installer from penguin.ooo/download: macOS 11+ (dmg), Windows 10+ (exe), or Linux x64 (AppImage, deb). The application is already logged in when opened, with no terminal required.
- Then run the command penguin web. It will http://127.0.0.1:7364 Provides a Web App and prints the first login link before setting a password; models are configured on the Models page. The desktop app and CLI share ~/.penguin/data.
- The image serves on port 7364, with data in the /data volume, supporting linux/amd64 and linux/arm64; latest follows changes to the image runtime or build inputs on main, and each release also has a version tag. The example listens only on the loopback address. For network-facing deployments, Compose, and upgrades, see the Docker quickstart.
- Each GitHub Release attaches exactly one package per target—x64 / arm64 for Linux and macOS, x64 for Windows, plus a runtime-free universal package. The same file serves both online and offline installation. Each package encapsulates the program payload, its SHA256 checksum, and the platform installer. Download the single file on a machine with network access, copy it to the target machine, extract it once, and run the bundled installer—no network is required, and there is no need to carry a separate checksum file (the sealed SHA256 is always verified).
- Linux (use penguin-linux-arm64.tar.gz on arm64); macOS (the example is Apple silicon; use penguin-darwin-x64.tar.gz on Intel); Windows (after extracting, double-click install.cmd, or run it in PowerShell).
- Chrome extension (optional): PenguinHarness Browser lets agents drive the tabs given to them in the user's own Chrome.
The same engine is scriptable and designed to be driven by agents. The development roadmap mentions: public release of the benchmark suite, Windows support, agent companies and templates, company-level self-evolution, and OpenShell integration (a shell with permission governance).
When developing this repository with a coding agent, its development Skills are located in .agents/skills/ (penguin-harness-dev, penguin-harness-frontend, penguin-harness-manual-test). Claude Code loads them without installation via the .claude → .agents symlink; other agents can link or copy .agents/skills/* into the directory where they read Skills. See CONTRIBUTING.md for the complete workspace guide, including development commands, quality gates, repository layout, and changelog rules.
Citations and acknowledgments: If you use PenguinHarness in research, please cite the relevant literature. The repository benefits from MinGit (a bundled POSIX shell and Git on Windows) and GenericAgent (built-in browser automation); see THIRD-PARTY-NOTICES.md for licenses. The project uses the Apache-2.0 license, © 2026 Prism Shadow. The maintenance team includes the organizations PrismShadow AI and Qinghe Dynamic, as well as Yaowei Zheng (LlamaFactory, PrismShadow), Leijun Zhou, Zhihao Liu, Xiang Qu, Hengjin Dou (all PrismShadow), Ning Wang, and Ming Lu (both Qinghe Dynamic).
Limitations: Apart from the above, more information about specific charges, prices, platform availability, performance metrics, or usage steps is not currently provided.
