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
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:
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.
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.
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.
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.
| 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.
| 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 |
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.
| 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.
# 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/cliThen:
penguin web # opens the same interface as the desktop app, in your browser
penguin chat # or talk to an agent in the terminalThe data root is ~/.adelie
8000
/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 --yesOn 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.
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.
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…
Thanks to everyone who has contributed to PenguinHarness — this fork stands on their code.
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}
}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.
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.