Skip to content

PenguinHarness 0.1.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 21 Jul 10:02
Immutable release. Only release title and notes can be modified.
v0.1.0
0e134e9
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

PenguinHarness 0.1.0 — the first feature release. An open-source agent harness you can self-host: a desktop-grade Web UI, a CLI, and a skill library an agent uses to evaluate and improve itself.

Install

curl -fsSL https://penguin.ooo/install.sh | sh
penguin web

Linux and macOS, x64 and arm64, with a bundled Node runtime — unpack and run. Or via npm (needs Node >= 24):

npm install -g @prismshadow/penguin-cli

Sign in with the seeded admin account admin / penguin-2026, then change the password.

Highlights

Agents that build agents. One sentence in, and an agent delivers a complete agent application — scaffold, code and run instructions. The bundled skills cover the whole path, including a RAG recipe that goes from a docs URL to a running, citation-backed Q&A app.

Recursive self-improvement. An Optimizer runs benchmarks against a target agent, reads the scores and traces to find where points were lost, and ships version N+1 — with a snapshot before every round and every request replayable in the Trace view.

Bring any model. 1000+ models through one gateway, with preset provider groups for DeepSeek, Anthropic, OpenAI, Google, Moonshot, Z.AI, plus the OpenRouter, SiliconFlow, Fireworks AI and Qwen gateways. Any OpenAI-compatible endpoint works. Credentials stay in a hidden project file and never enter the model context.

Benchmarks. On a 15-task data-analysis suite, PenguinHarness on DeepSeek V4 Pro reaches the best accuracy of the three harnesses tested (66.67% against 53.33% for both Claude Code and OpenAI Codex) at 1/70 of Claude Code's cost. On 40 coding tasks it ties Codex at 71.25% and trails Claude Code's 86.25%, for $3.81 against $220.08 and $146.97. Full tables at https://penguin.ooo/#benchmark

Web App. Multi-session chat with streaming output, tool approval and image paste; a workspace file browser; the Trace view; a cost centre; an evaluation centre; and a skill library with one-click installs and update reminders.

Notable in this release

  • A model reference is always an explicit (provider, model_id) pair. The provider is never inferred, guessed or defaulted. Gateways resell vendor models under the vendor's own id, so guessing a group could send one vendor's API key to another vendor's endpoint. --provider is now required wherever a model id is given.
  • AgentHub 0.4.0. Message content carries an opaque fidelity payload, replacing the item-level signature/phase fields and preserving provider wire data verbatim through Trace, replay and resume.
  • Language and theme are shared between the landing page and the docs site.

Requirements

Linux or macOS (x64 / arm64). The installer bundles its own Node runtime; installing from npm needs Node >= 24. All data stays under ~/.penguin/data.

Full detail: changelog/0.1.0/

Assets 15
Loading