Prism-Shadow/penguin-harness

What it does

PenguinHarness is a framework for recursive self-improvement (RSI) of agents. The underlying model stays fixed; what evolves is the harness around it — prompts, Skills, tools and hooks. A user picks an algorithm and a benchmark, and an agent improves another agent round by round, with each candidate snapshotted before changes and scored in an Evaluation Center.

Nine RSI algorithms ship as plugins (default, OPRO, APE, ACE, AWM, SkillAdam, SkillOpt, OpenSkill, TRACE), all running on the same built-in benchmarks and scoreboard. New projects include seventeen Docker-based benchmarks drawn from rag-bench-essential, DeepSWE, AutomationBench, Terminal-Bench-Science, Terminal-Bench and GDPEVO Essential, with more in the separate benchmark repository. The product also covers app generation from one sentence, browser automation, hot-reloaded workbenches, SSH remote machines and a streaming Agent API.

It suits developers and researchers who want to experiment with agent self-evolution and evaluation, and users who want a ready-made agent app platform.

Getting started

Desktop users can download an installer for macOS 11+, Windows 10+ or Linux x64 from the download page; the app needs no terminal.

For CLI or server use (Node >= 24):

npm install -g @prismshadow/penguin-cli
penguin web

penguin web serves the Web App at http://127.0.0.1:7364 and prints a first-login link until a password is set. Configure models on the Models page. The desktop app and CLI share data under ~/.penguin/data.

Docker is also supported:

docker run -d --name penguin -p 127.0.0.1:7364:7364 -v penguin-data:/data hiyouga/penguinharness:0.2.18
docker logs penguin   # read the first-login link

After setup, add a model and run a task:

penguin config model add --provider deepseek --model-id deepseek-flash --api-key sk-... --set-default
penguin run -m "Create hello.txt containing Hello, Penguin"

Offline packages for air-gapped machines are attached to GitHub Releases; extract the one matching your platform and run the bundled installer. For the SDK, @prismshadow/penguin-core exposes createAgent, sessions and streaming outputs, as shown in the README's TypeScript sample. Full setup details are in the documentation.

Constraints

• The CLI and Docker server require Node >= 24 or Docker; the desktop app targets macOS 11+, Windows 10+ and Linux x64.
• Model access is bring-your-own-key through the Models page or penguin config; supported providers include OpenAI, Anthropic, Google Gemini, DeepSeek, Moonshot, Z.AI and OpenRouter-compatible endpoints.
• The example Docker command binds to loopback only; exposing it on a network requires additional configuration.
• Agents execute tools and shell commands, so tool-call approval is part of the runtime design; plan permissions accordingly.
• Benchmarks run as Harbor tasks in Docker, so benchmarking adds Docker to the requirements.
• Licensed under Apache-2.0.

HotRepo's take

No HotRepo ranking or snapshot history is available for this repository, so there is no station-side trend data to weigh. What stands out technically is the breadth of the design: nine published RSI algorithms share one benchmark and snapshot pipeline, and the same engine is exposed through a desktop app, Web App, CLI and SDK. The project's own results describe leading accuracy on a data-analysis suite at a fraction of rival cost, but those figures are author-reported and benchmark-specific; they should be read as the project's claims rather than independently verified outcomes. The presence of a maintained benchmark repository and a public roadmap for company-level self-evolution and OpenShell integration suggests active work beyond the README.