AI-Native Workspace
RedDragonHQ/ai-native-workspace
SO pessoal baseado em camadas para agentes de IA com swapping de modelo/agente
Ver no GitHubFramework em camadas para transformar qualquer agente de IA em um sistema operacional pessoal. Funciona com Claude Code, Codex, Cursor ou qualquer agente que leia markdown, desacoplando modelo, agente e protocolo para máxima flexibilidade e portabilidade.
Instalação
npx convez-ai RedDragonHQ/ai-native-workspace1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Construir workspaces onde agentes IA operam como operadores primários
- Manter continuidade entre diferentes modelos (Claude, DeepSeek, Qwen) sem reescrever workflows
- Implementar protocolo multi-agente (Claude Code, Codex, Cursor) sobre mesma base
- Criar fallback humano estruturado quando agentes falham ou APIs descem
- Organizar pesquisa/engenharia com estado persistido em CLAUDE.md entre sessões
Veredito

Framework bem-pensado que resolve o problema real de lock-in em ecosistemas de agentes IA — elegante na separação de camadas (modelo/agente/protocolo). Exige maturidade arquitetural para extrair valor máximo, mas é genuinamente inovador na premissa de workspace como OS. Recomendado para equipes que planejam além de um agente único.
README
A tradução para o português deste README ainda está na fila.
AI-Native Workspace
A layered framework for turning any AI coding agent into a personal operating system.
Works with Claude Code, Codex, Cursor, or any agent that can read a markdown file.
🌏 Non-English readers: Chinese versions of every document live in zh/ — zh/README.md · zh/ARCHITECTURE.md · zh/BOOTSTRAP.md · zh/examples/sample-workspace/.
This is not another .cursorrules template. It's an opinionated architecture for running a real research/engineering workspace where the AI agent is the primary operator — and humans are the fallback when agents aren't available.
Why this framework
Most "Claude Code workflow" repos fall into one of two camps:
- Dotfiles dumps — my
.claude/config, my rules, my prompts. Hard to generalize. - Single-ecosystem lock-in — designed for one agent + one model + one provider. Breaks the day you want to switch.
This framework is built on a different premise:
The workspace is the OS. The agent is the process. The model is the kernel.
You can swap any layer without rewriting the others.
Multi-layer decoupling — the actual differentiator
| Layer | Single-ecosystem approach | This framework |
|---|---|---|
| Model | Claude-only (or GPT-only) | DeepSeek, Claude, Qwen, OpenAI, Gemini — same .env swap |
| Agent | Claude Code only | Claude Code, Codex CLI, Cursor, custom LangGraph agent |
| Protocol | Provider-specific (Anthropic MCP, OpenAI function calling) | File-based CLAUDE.md + CLI + MCP (optional) |
| Application | Vendor templates | Your projects, your PKM, your workflow |
Why this matters: if Anthropic ships a better agent next year, you don't rewrite your workspace. If DeepSeek undercuts Claude on price, you don't rewrite your workflow. If you outgrow MCP, you don't lose your knowledge base.
The framework is protocol-based, not product-based.
AI-first, human-readable
CLAUDE.md is optimized for agent consumption — structured so an AI agent can load it and start working immediately. But it's not agent-exclusive: the same file is fully human-readable, and when agents aren't available, a human can step into the agent layer and operate the workspace using the same protocols.
- Layered cognitive loading: global → workspace → project
- Progressive context: read the index first, dive into details on demand
- Structured continuity: cross-session state lives in CLAUDE.md and project docs, not in chat history
- Workflow contracts: session-end protocols, recovery protocols, audit trails
Humans don't just get side benefits — they're the fallback operator. When the agent layer fails (API down, model degraded, edge case the agent can't handle), the human reads the same CLAUDE.md and picks up where the agent left off.
The 4-layer architecture
┌─────────────────────────────────────────────────────────┐
│ Application │
│ Projects · Inbox (draft buffer) · Wiki (matured KB) │
├─────────────────────────────────────────────────────────┤
│ Protocol │
│ CLAUDE.md conventions · CLI tools · MCP Servers │
├─────────────────────────────────────────────────────────┤
│ Agent │
│ Claude Code · Codex · Cursor · any markdown-aware agent│
├─────────────────────────────────────────────────────────┤
│ Model │
│ DeepSeek · Claude · Qwen · OpenAI · local │
└─────────────────────────────────────────────────────────┘
Each layer has a single responsibility. The boundaries are enforced by file conventions, not code — which is why any agent that can read markdown can run this framework.
See ARCHITECTURE.md for the full design.
What's in this repo
ai-native-workspace/
├── README.md ← You are here (English)
├── zh/README.md ← 中文版
├── ARCHITECTURE.md ← 4-layer design, trade-offs, anti-patterns (English)
├── zh/ARCHITECTURE.md ← 中文版
├── BOOTSTRAP.md ← Agent init instructions (English)
├── zh/BOOTSTRAP.md ← 中文版 (agent 初始化指令)
├── examples/
│ └── sample-workspace/ ← Minimal runnable workspace (English)
├── zh/examples/
│ └── sample-workspace/ ← 中文版
└── LICENSE
BOOTSTRAP.md is the interesting file. It's the document an AI agent reads to set up a fresh workspace from scratch. Paste it into Claude Code / Codex / Cursor in an empty directory and the agent will build the structure for you. Chinese users: use zh/BOOTSTRAP.md instead.
Quick start
Option A: let an agent bootstrap it
- Create a new directory for your workspace.
- Open Claude Code / Codex / Cursor in that directory.
- Paste the contents of
BOOTSTRAP.mdas your first prompt. - The agent will scan your directory and build the structure.
Option B: copy the example
cp -r examples/sample-workspace ~/my-workspace
cd ~/my-workspace
$EDITOR CLAUDE.md # adjust to your context
Chinese version:
cp -r zh/examples/sample-workspace ~/my-workspace
cd ~/my-workspace
$EDITOR CLAUDE.md
Then what?
Start using it. The framework reveals itself through use:
- Day 1: you write
inbox/YYYY-MM-DD-thing.mdat end of session - Week 2:
inbox/TODO.mdbecomes your cross-session state - Month 2:
wiki/emerges when inbox gets crowded - Month 3+: CLAUDE.md and project docs accumulate persistent facts about your preferences
No setup ceremony. The architecture is the daily practice, not the installation.
Key ideas (TL;DR)
- Inbox is the buffer. Low-friction capture. Anything can land in
inbox/— you sort later. - Wiki is the sediment. When inbox items get referenced 3+ times, they graduate to
wiki/. - MANIFEST is the index. Don't read the whole wiki. Read MANIFEST.md (one-line summaries), then dive.
- Anti-patterns are the guardrails. The framework is defined as much by what it forbids as what it enables.
Anti-patterns (a taste)
| ❌ Don't | ✅ Do |
|---|---|
| Start with multi-agent | Single agent; split only when context is genuinely exhausted |
| Load all workspace files at boot | Read the index, drill into what you need |
| Skip inbox and write straight to wiki | Inbox is the buffer; let sediment form naturally |
| Write a 500-line CLAUDE.md | ≤ 100 lines; details go in project CLAUDE.md |
| Treat TODO.md as a command queue | It's a draft — validate before acting |
| Auto-write inbox files without confirmation | Prompt the user; don't surprise them |
See ARCHITECTURE.md for the full list with rationale.
When to use this
Use it if:
- You work on multiple projects with an AI coding agent and context loss is a real problem
- You've tried
.cursorrules/.claude/CLAUDE.mdbut they don't scale past one project - You want a framework that survives switching models/agents/providers
- You're doing research or infrastructure work where knowledge accumulates across months
Don't use it if:
- You're working on a single short-term project (a simple CLAUDE.md is enough)
- You don't need cross-session memory (just use chat history)
- You're not using an AI agent daily (the framework rewards consistent practice)
Origins
This framework was extracted from a real research workspace running 8+ projects (ML research, multi-agent systems, infrastructure, security research, PKM) over 6+ months of daily use. It was refined through repeated failures — the anti-patterns list is longer than the features list because each one represents a real mistake.
The workspace is called RedDragon Workspace internally. This release strips the personal content and keeps the architecture.
Contributing
Issues and PRs welcome. Especially:
- Adapters for other agents (Codex, Aider, Continue, etc.)
- Example workspaces for specific domains (research, SWE, content)
- Translations of BOOTSTRAP.md and ARCHITECTURE.md
License
MIT