A local-first multi-agent app development platform, built on PenguinHarness
Adelie
A local-first multi-agent app development platform, built on PenguinHarness
[!IMPORTANT] Adelie is a fork of PenguinHarness. The code tree, the engine, the Web frontend and the protocol all come from PenguinHarness (Apache-2.0); Adelie is the productised line built on top of it — its own brand and interface, its own data root and ports, its own release pipeline.
- Where it came from, what the licence requires, what we changed:
FORK.md- How far the work has got:
FORK-PROGRESS.md- The upstream project's own channels — repository, website, docs, blog, community, Discord, X, WeChat, Product Hunt — are theirs, not Adelie's: https://github.com/Prism-Shadow/penguin-harness · https://penguin.ooo/
- The PenguinHarness and PrismShadow names and marks belong to upstream. Adelie does not use them as its own name, icon or domain.
简体中文 | English
Why Adelie
Build agents by hand with LangChain at 1× speed; build agents with Adelie at 100× speed.
Adelie runs on your own computer or server and strings together the creation, evaluation, optimisation and deployment of agent applications. Three reasons, in ascending order:
1. 🏆 Excellent results at a fraction of the cost
A deliberately small tool set over a clean interface: fewer tool calls, fewer tokens, and deep adaptation to open models such as DeepSeek. Same tasks, each on its usual model, head to head:
The most accurate on the data-analysis set — at 1/70th of Claude Code's cost.
2. ⚡ One sentence in, a runnable agent app out
Describe what you want in one sentence and Adelie builds the whole agent application — scaffold, code, run instructions, in one go:
Collect the docs at https://github.com/ericbuess/claude-code-docs and build a RAG Q&A app that answers as a Claude Code configuration expert, with cited sources.
This is what came out — a docs expert: retrieval-augmented, citations that open the original file, sample questions built in:
And generating the whole RAG app cost ¥0.2 ($0.02) in tokens — on the DeepSeek V4 Pro model.
3. 🧬 A native agent self-evolution engine
With Adelie's skill library an agent evaluates and optimises itself: run a Benchmark, find where the score went, ship version N+1 — with a snapshot before every round and every request replayable in the trace viewer.
About the images in this README: the benchmark chart comes from the PenguinHarness authors' measurement of the very codebase this fork is built on — Adelie runs the same engine and interface, so what it shows is this fork's capability too. Every screenshot is Adelie's own interface, captured from a local instance.
Built-in plugin library
Four kinds of plugin ship in the box (documented under packages/docs) — Skills,
plus the session hooks that drive goal mode and continual learning; an agent can also write and
optimise its own Skills:
| Category | Plugins |
|---|---|
| Office Productivity | data-analysis, use-firecrawl, use-offerdao, browser-automation, use-bento-slides, use-lieflat-charts, use-wps-office, use-seedance, use-seedream, humanizer, goal, continual-learning, csu-mail, lesson-video, requirements-box |
| Software Development | software-development, use-claude-code, wechat-miniprogram, gpu-model-serving, web-vscode |
| AI App Development | agent-development, model-development, skill-porting, agent-tuning |
| Agent Company | agent-company |
The plugin market page lists every plugin that ships with the build, each with its own README, version, licence and keywords, and installs one in a click:
The desktop app also carries a built-in browser in its side dock. An agent drives it through
penguin browser and the browser-automation plugin: reading pages, clicking and typing, and
extracting data such as Amazon orders — signed in with credentials imported from your own browser.
Supported models
| Model | Providers |
|---|---|
| DeepSeek V4 | DeepSeek, OpenRouter, Fireworks AI, SiliconFlow, TokenDance, Qwen Token Plan, Qwen Pay-As-You-Go |
| Kimi K3 | Moonshot AI, OpenRouter, Fireworks AI, TokenDance, Qwen Pay-As-You-Go |
| GLM 5.3 | Z.AI, OpenRouter, TokenDance |
| Hunyuan 3 | OpenRouter |
| Qwen 3.8 Max | Qwen Token Plan, Qwen Pay-As-You-Go, OpenRouter, TokenDance |
| GPT 5.6 | OpenAI, OpenRouter |
| Gemini 3.7 Flash | Google Gemini, OpenRouter |
| Claude 5 | Anthropic, OpenRouter |
| Inkling | OpenRouter, Fireworks AI |
The table lists only the newest generation of each family; the Models page carries the full built-in catalog. Anything speaking the OpenAI protocol works: pick a preset, or connect 1000+ hosted and local models with a custom endpoint.
System requirements
| Requirement | Support |
|---|---|
| Operating system | Windows 10+ (x64), Linux (x64), macOS (arm64 / x64) |
| Runtime | The installers bundle Node; npm and from-source installs need Node >= 24 |
| Model | an API key for at least one model provider |
Install
Adelie ships its own artifacts — desktop installers, CLI bundles and npm packages, every release on GitHub Releases. That is the only place they are published: the Alibaba Cloud OSS mirror this fork used to keep was retired, so there is one address to check and it is always current. A release is always a whole version: the desktop app, the CLI, the server and the Web frontend are the same number.
Desktop app
| Platform | Artifact |
|---|---|
| Windows 10+ (x64) | adelie-desktop-win32-x64.exe — NSIS installer, per-user, installs under %LOCALAPPDATA% |
| Linux (x64) | adelie-desktop-linux-x86_64.AppImage or adelie-desktop-linux-amd64.deb |
| macOS | not built yet — the dmg/zip targets wait for an Apple Developer ID |
The Windows installer is not code-signed yet, so SmartScreen shows its warning once ("More info → Run anyway"). The app checks for updates itself against this repo's GitHub Releases, so "Check for updates" inside the app is the normal way to upgrade.
Command line and Web App
# Linux / macOS, one line: installs penguin under ~/.adelie, with its own Node runtime
curl -fsSL https://github.com/lmliheng/Adelie/releases/latest/download/install.sh | sh
# or via npm (any platform, needs Node >= 24 on the machine)
npm install -g @adelie/cli
Then:
penguin web # opens the same interface as the desktop app, in your browser
penguin chat # or talk to an agent in the terminal
The data root is ~/.adelie/data (ADELIE_HOME overrides it; the older PENGUIN_HOME is still
honoured, so an install that predates the rename keeps reading its own data). Upgrading replaces
bin/, lib/, web/ and node/ and never touches the data root:
penguin update --check
penguin update --yes
On first start the server prints a first-login link; open it to claim the built-in admin account and
set its password (there is no initial password to type). Configure a model on the Models page —
a new instance needs one API key before it can run a Task.
[!WARNING] The upstream project's own ways in —
curl https://penguin.ooo/install.sh | sh,npm install -g @prismshadow/penguin-cli,docker run hiyouga/penguinharness, and the desktop downloads on https://penguin.ooo/download — install PenguinHarness, not Adelie. Adelie has its own installers, its own npm scope and its own update feed; use the ones above.
The command name is still penguin and the packages still carry the penguin- stem (Adelie's own
scope, @adelie/*); the rename is being done in stages and what is left of it is tracked in
FORK-PROGRESS.md.
Contributing
pnpm install && pnpm build # build first: core's exports point at dist/
pnpm dev # server + Web frontend together (prefixed logs, dependencies built once)
The full workspace guide is CONTRIBUTING.zh.md (that file is upstream's
and remains broadly accurate): development commands, quality gates, repository layout, changelog
rules. The upstream marketing site, packages/landing, has been deleted from this fork.
This fork's own docs are collected in docs/ — the fork story, the progress log and
every release note — so the repository root stays down to the conventional files.
Upstream roadmap
This is the upstream project's roadmap, kept here to show where the base is heading. Adelie's own
plans live in FORK-PROGRESS.md.
- Benchmark suite released
- Desktop app
- Windows support
- Agent company and templates
- Company-level self-evolution
- OpenShell integration (a shell with permission control)
- More to come…
Upstream contributors
Thanks to everyone who has contributed to PenguinHarness — this fork stands on their code.
Citation
If you use this code, please cite the upstream project (Adelie is its fork):
@software{penguinharness2026,
author = {{PrismShadow Team}},
title = {PenguinHarness: Efficient Self-Improving Harness for Everyone},
year = {2026},
url = {https://github.com/Prism-Shadow/penguin-harness},
license = {Apache-2.0}
}
Acknowledgements
This project benefits from:
- MinGit: the bundled POSIX shell and Git on Windows
- GenericAgent: built-in browser automation
See THIRD-PARTY-NOTICES.md for the full list.
License
Apache-2.0 © 2026 Prism Shadow — that is upstream's copyright. Adelie is its fork and is
released under the same licence; see FORK.md.
Upstream PenguinHarness is built with ❤️ by Yaowei Zheng, the author of LlamaFactory, PrismShadow AI Team and Fable 5. Adelie is a fork of that work.