A local-first multi-agent app development platform, built on PenguinHarness
Built-in plugins
The Adelie plugin library. Each plugin is its own npm package — @adelie/<name>, a directory under plugins/ — with a plugin.json manifest (and an icon.svg beside it: the icon of everything the plugin ships, which its installed skills and hook package inherit) plus the content it ships: skills (skills/<name>/SKILL.md, installed into an Agent's agent_state/skills/) and/or a hook package (hooks/*.mjs, installed into agent_state/hooks/<plugin>/ with a generated hooks.json). The loader lives in @adelie/core, which depends on these packages and reads their directories at runtime — the files are the source of truth.
Versions are dates with a sequence number — YYYY.MM.DD.N — on the manifest and on every skill a plugin ships (the manifest's dated version is the plugin's own; the package's npm version follows the release). Skills follow the "index first, body on demand" design: only their metadata is injected into an Agent's system prompt; the Agent reads the full SKILL.md via shell when it actually needs it. Hook scripts are plain Node (builtins only): the harness runs them as subprocesses at the loop's hook points with { hook, session_id, trace_path } on stdin and reads their JSON answer from stdout.
Included plugins, by category (PLUGIN_CATEGORIES in packages/core/src/plugins/index.ts; a plugin with no or an unknown category lands in an "Other" group). A plugin built around someone else's product carries a use- prefix — use-firecrawl, use-bento-slides, use-claude-code — so the package name says what it is for rather than claiming the product; the skills inside keep their own names:
| 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 |
humanizer, use-claude-code, continual-learning, agent-company, csu-mail, wechat-miniprogram, lesson-video, requirements-box, gpu-model-serving and use-lieflat-charts carry preinstall: false, so default_agent does not get them at initialization — they are installed from the library on demand (agent-company by the organization itself, when it creates the CEO and hires employees, or by hand onto any Agent that should be able to create one). goal is the stop hook behind goal mode: its start.mjs, a user_prompt command marked "trigger": "host" so that it runs only when a goal starts, writes the Session's GOAL.json and composes round 1; its stop.mjs reads the Trace after every Task and injects the next round or ends the goal. continual-learning hands a long task's condensed excerpt to a background subagent that folds the findings into the agent's skills.
browser-automation drives the desktop app's built-in browser through penguin browser: GenericAgent's scan → act → check loop over simplified HTML, trusted clicks and typing, and sign-ins imported from the system browser, with Amazon orders as the worked example.
use-lieflat-charts turns data into HTML charts, and — only when the user asks for it — into a paged HTML/PDF deck or a full-page report. It is a port of the upstream lieflat-charts skill by 「躺在废墟里」, and the skill body, the chart catalogues, the templates and the examples are that project's own text, shipped unchanged: the plugin only adds the attribution, the ## Before you start section every library skill carries, and the relative-path fixes for the upstream preview images it does not ship. What makes it work is the discipline, not the templates alone — every chart starts from a real template implementation (Lupi editorial, Basics, Glance, Maps, an interactive big graph), keeps one colour system per deliverable (Mono by default, three built-in presets, or a custom palette the user supplies), and inherits the same typography, hairlines, whitespace and entrance animation. It ships about 2.5 MB of templates, so it is installed onto the Agents that draw rather than onto everyone. Its licence is the upstream PolyForm Noncommercial 1.0.0, carried in LICENSE with the provenance in NOTICE.md — noncommercial use, modification and redistribution are free; commercial use needs the author's permission.
agent-tuning powers the self-improvement loop: create the Target Agent, design a Benchmark, evaluate it, optimize it to version N+1 with a snapshot before every round.
gpu-model-serving stands up an OpenAI-compatible endpoint on a rented or owned GPU box: pick the instance and where the weights live, take an already-quantized checkpoint, build the pip CUDA toolchain, start serve.sh detached, then put the service through the five checks that decide it is done. Its reference/pitfalls.md records what the first attempt costs — symptom, cause, fix — and reference/verify.md holds the commands and what each answer should look like. Installed onto the Agents that serve models rather than onto everyone.
agent-company is the whole toolkit of company mode, entry point included. company-setup is that entry point: the skill that creates an organization with the user, asking one question at a time and ending at penguin org create, leaving hiring and tickets to the CEO. company-employee is the protocol every desk and ticket session follows — including the rule that whatever touches the user's machine, spends money or reaches outside the organization is asked of the board in the all-hands channel first — and company-ceo, company-hr and company-finance are the playbooks of those titles. company-research is what a research organization's authors and reviewers add: fix the harness and the metric, run the experiment loop inside a resource envelope the board approved, keep only improvements, and put every claim through a reviewer who tries to break it. company-mirror is the seventh: the protocol a company of digital twins runs on instead, where every employee mirrors a real colleague and the company relays between people rather than filing tickets. Because the plugin is preinstall: false, company-setup is no longer on every Agent by default — an Agent gets it by installing agent-company from the library (the Agents page), and the CEO and the employees carry it already.
Documentation
Development
The plugins are data, not code — there is nothing to build here. The loader, its types and their tests live in packages/core:
pnpm --filter @adelie/core build # includes the plugin loader
pnpm --filter @adelie/core test # loader, README tables, the hook scripts against fake Traces