Digital FTE: Funcionário IA Autônomo
rj41-w2/Personal-AI-Employee-Hackathon-0
FTE autônomo local com aprovação humana e segurança em camadas
Ver no GitHubUm servidor MCP que implementa um funcionário IA autônomo local-first, com orquestração de tarefas via Obsidian, watchers para Gmail/Odoo/LinkedIn e execução de ações com aprovação humana. Combina Claude Code, Python e dashboard Next.js em um fluxo perception → reasoning → action.
Instalação
npx convez-ai rj41-w2/Personal-AI-Employee-Hackathon-01 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
- Automação de tarefas repetitivas com revisão humana
- Orquestração de fluxos multi-canal (email, CRM, redes sociais)
- Agente de produtividade pessoal com controle de segurança
Veredito

Projeto ambicioso e bem-estruturado para automação pessoal com controles de segurança robustos (aprovação humana, flags de ambiente). Exige Python/Node.js avançado e setup local complexo; ideal para quem quer full control sobre automação, mas não é plug-and-play.
README
A tradução para o português deste README ainda está na fila.
Digital FTE: Personal AI Employee
Your life and business on autopilot: local-first, agent-driven, and human-in-the-loop.
This project is an autonomous "Digital FTE" built for the Personal AI Employee Hackathon. It combines an Obsidian vault, Python watchers, local LLM drafting, MCP action servers, and a Next.js dashboard.
Architecture
The system runs as a perception -> reasoning -> action loop:
- Watchers monitor Gmail, Odoo, LinkedIn, and local files.
- Watchers create Markdown tasks in
AI_Employee_Vault/Needs_Action. - The orchestrator drafts plans and actions into
Pending_Approval. - A human reviews the draft and moves it into
Approved. - The MCP executor validates approval metadata and environment safety flags before any live action runs.
Project Structure
.
|-- AI_Employee_Vault/ # Obsidian vault and workflow folders
| |-- Dashboard.md # Real-time status updates
| |-- Needs_Action/ # New tasks waiting for AI drafting
| |-- Pending_Approval/ # Drafts waiting for human review
| |-- Approved/ # Human-approved tasks
| `-- Done/ # Successfully completed tasks
|-- ai-dashboard/ # Next.js dashboard
|-- src/
| |-- orchestrator.py # Main workflow controller
| |-- api_server.py # FastAPI backend for the dashboard
| |-- mcp/ # Email, LinkedIn, and Odoo MCP servers
| |-- skills/ # Drafters and MCP executor
| |-- watchers/ # Gmail, Odoo, filesystem, and LinkedIn watchers
| `-- prompts/ # Prompt rules and guardrails
|-- GEMINI.md # Agent instruction context
`-- main.py
Getting Started
Prerequisites
- Python 3.11 or newer
- Node.js 20.9 or newer
- Obsidian pointed at
AI_Employee_Vault - Ollama running locally if you use the built-in local LLM drafting flow
Installation
git clone https://github.com/rj41-w2/Personal-AI-Employee-Hackathon-0
cd Personal-AI-Employee-Hackathon-0
python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt
cd ai-dashboard
npm install
cd ..
copy .env.example in .env
Fill in .env with your local credentials. Never commit real secrets.
Usage
Start the orchestrator:
python src/orchestrator.py
Start watchers in separate terminals:
python src/watchers/gmail_watcher.py
python src/watchers/filesystem_watcher.py
python src/watchers/odoo_watcher.py
Start the dashboard:
python src/api_server.py
cd ai-dashboard
npm run dev
Live Action Safety
Moving a file to AI_Employee_Vault/Approved is not enough to send an email, post to LinkedIn, or write to Odoo. Live execution is blocked unless all required safety checks pass.
To enable live actions, set:
ENABLE_LIVE_ACTIONS=true
Then add approval metadata to the approved Markdown file:
Approval: approved
Approved_By: your-name
Additional per-action guards:
- Email sending requires
EMAIL_ALLOWED_RECIPIENTSorEMAIL_ALLOWED_DOMAINS. - LinkedIn posting requires
LINKEDIN_POSTING_ENABLED=true. - Odoo write actions require
ODOO_WRITE_ACTIONS_ENABLED=true. - Odoo read actions require
ODOO_READ_ACTIONS_ENABLED=true.
With live actions disabled, approved files are treated as dry runs and no external MCP tool is called.
Security & Privacy
- Local-first by default.
- Real secrets belong in
.env, not in Git. - Gmail OAuth files (
credentials.json,token.json) are ignored by.gitignore. - Review
SECURITY.mdbefore publishing or deploying.
License
MIT. See LICENSE.