PenguinHarness

PenguinHarness is an open-source AI agent harness built by the PrismShadow AI Team, led by Yaowei Zheng (author of LlamaFactory). It runs fully locally — all data stays in ~/.penguin/data — and reaches 1000+ online and local models through a unified model gateway. The project is licensed under Apache-2.0 and available on GitHub.

What It Is

PenguinHarness is a complete TypeScript stack for constructing and evolving AI agents. It ships four tightly integrated layers — an SDK (@prismshadow/penguin-core), a CLI (@prismshadow/penguin-cli), a server (@prismshadow/penguin-server), and a browser web app (@prismshadow/penguin-web) — all sharing a single data directory and a unified message protocol called OmniMessage. The project describes itself as "the first open-source harness to ship 'agents building agents' and recursive self-improvement."

Three Core Pillars

The harness is organized around three design principles:

• Simplest Is the Best — A deliberately minimal toolset over clean low-level interfaces: dedicated file tools (read_file, edit_file, write_file) and a shell fallback (exec_command), tuned to minimize token usage on open models like DeepSeek.
• Harness for Building Agents — With the PenguinHarness SDK, an agent builds complete agent applications autonomously from a single sentence of input, delivering scaffold, code, and run instructions end to end.
• Harness for Recursive Self-Improvement — With PenguinHarness Skills, an Optimizer orchestrates multiple Evaluators to score a Target Agent in parallel, uses scores and run traces to find where points were lost, and upgrades the agent from version N to N+1 — with a snapshot before every round.

Architecture and Design Tenets

The harness enforces a CONTRACT.md that defines the boundary of evolution: self-improvement is strictly confined to Workspace and Skills, while the harness kernel and its safety mechanisms never change. Key design tenets include:

• Full tracing — Every model request and tool call is written to the Trace in full, including token count, latency, and failure reason; sessions are fully restorable from the Trace.
• Approvals and audit — Every tool call requires user approval before it runs, and every decision leaves an audit record.
• Credential isolation — API keys land as hidden 0600 files, are barred from the system prompt, and stay masked throughout the UI.
• Model decoupling — Models are not bound to agents; you pick one per session and can switch without rewriting the agent.
• Progressive loading — Content is indexed first and read on demand, never dumped wholesale into context.
• Error convergence — Errors split into retryable and fatal; retryable ones retry automatically, fatal ones become messages the model can react to.

Built-in Skill Library

Four skill groups ship out of the box, and agents can write and optimize their own:

• Office Productivity: data-analysis, firecrawl
• Software Development: web-design, software-engineering
• AI App Development: penguin-sdk, penguin-cli, agenthub-models, vllm, ollama, llamafactory
• Agent Tuning: agent-creation, benchmark-design, agent-evaluation, agent-optimization

Deployment and Platform Support

PenguinHarness supports two installation paths that share the same local data root:

• Desktop app — A double-click installer for macOS 11+, Windows 10+, and Linux (AppImage/deb) that embeds the server and opens already signed in.
• CLI — A one-line installer (curl | sh on Linux/macOS, PowerShell on Windows) or npm install -g @prismshadow/penguin-cli; penguin web then serves the full web UI at http://127.0.0.1:7364. Offline/air-gapped installs are also supported via GitHub Release packages.

The system requires only a single CPU at minimum and supports x64 and arm64 architectures. Node >= 24 is required for npm installs; the one-line installer bundles its own runtime.

Update: v0.2.1

The latest release is v0.2.1, published on 2026-08-04. The repository was created on 2026-07-19 and last pushed on 2026-08-07, indicating rapid early development. Recent blog posts announce availability of Kimi K3 and free models like Ling 3.0 Flash in PenguinHarness, as well as a Fireworks AI AMD Developer Program credit offer. The roadmap lists upcoming items including a public benchmark suite release, agent company templates, company-level self-evolving, and OpenShell integration.