Agentic OS
Taran132g/agentic-os
Agente IA autônomo local com multi-workers, auto-correção e integração Telegram
Ver no GitHubSistema autônomo de agentes IA pessoais que executa tarefas end-to-end (pesquisa, candidaturas, organização) via dashboard web ou Telegram. Integra Claude Code CLI com correção automática, controle de computador Mac e persistência em Obsidian.
Instalação
npx convez-ai Taran132g/agentic-os1 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 candidaturas a vagas com tailoring de currículo
- Orquestração de tarefas autônomas com aprovação de ações sensíveis
- Pesquisa e organização pessoal com persistência em Obsidian
Veredito

Projeto maduro e bem-arquitetado para quem quer agente IA autônomo local com controle total. A combinação de multi-workers assíncronos, self-correction via leitura de código-fonte e integração Mac é sofisticada; exige Python e setup local significativo.
README
A tradução para o português deste README ainda está na fila.
Agentic OS
A personal autonomous AI agent system running locally on macOS. PAIS handles tasks end-to-end — research, job applications, personal organization — via a web dashboard or Telegram, with full computer use capabilities.
How It Works
- Submit a task via Telegram (
/task ...) or the web dashboard - PAIS routes it to the right agent: Career, Personal, or General
- The agent runs autonomously using Claude Code CLI
- Results are logged to your Obsidian vault; you're notified on Telegram
- Sensitive actions (sending emails, pushing to GitHub) require your explicit approval
Features
Multi-Agent Architecture
Four independent async workers run in parallel — career, personal, general, and a Telegram intake worker. Workers are freed immediately after a task completes; verification and self-correction run asynchronously so the queue never stalls.
Claude-Only LLM Runner
Runs Claude Code CLI exclusively. On usage limit errors, PAIS waits 30 seconds and retries automatically — no Gemini dependency, no manual intervention needed.
Self-Correction Loop
When a task doesn't complete correctly, click "Didn't complete correctly" and type what went wrong. PAIS reads its own source code, identifies the root cause, patches the file, writes a lesson to lessons.md, hot-reloads the module, and re-queues the original task — all without a restart.
Persistent Learning (lessons.md)
An append-only self-correction log stored at the project root. Every agent reads the most recent lessons before executing a task, so PAIS improves from past mistakes across sessions.
Computer Use
PAIS can control your Mac desktop: take screenshots, click, type, scroll, drag, run AppleScript, OCR the screen, and open apps. Useful for automating GUI-only workflows.
Career Workflow
Searches for jobs matching keywords, tailors resume bullets per job description, and uses Playwright to open application forms and pre-fill them — stopping before submit so you review before anything is sent. Search results are cached to disk so the Career Ops page restores them on page load.
Personal Tasks Workflow
Handles research, planning, and personal organization. Analyzes the task, creates a structured plan, and writes it to your Obsidian vault. Includes the "Didn't complete correctly" self-correction button.
Task Log
Every completed task is saved to the vault with a YYYY-MM-DD HH-MM filename and a ## Requested: field. The dashboard history view is sorted by modification time with full YYYY-MM-DD HH:MM timestamps so you can see exactly when each task ran.
Telegram Bot
/task <description>— queue a task/status— check queue and pending approvals- Inline buttons for routing (Career / Personal / General) and Approve/Deny on sensitive actions
Web Dashboard
Real-time FastAPI dashboard with WebSocket push. Shows live task output, task history from the vault, system logs, and manual controls. Includes a self-correction feedback panel and a dedicated Career Ops page.
Obsidian Integration
All completed tasks, career search results, and personal plans are written as Markdown notes to a local Obsidian vault for long-term memory.
Approval Gate
Any action with external side effects (sending email, pushing code, posting online) is paused and a request is sent to both Telegram and the dashboard. Nothing irreversible happens without your explicit approval.
Always-On Mac
PAIS runs via a macOS LaunchAgent wrapped with caffeinate -i -s so the machine never sleeps while the agent is active.
Task Log Audit (May 2026)
After the May 11 build session, a 7-task end-to-end audit was run to stress-test PAIS after the worker stall fix and self-correction loop were introduced. Each task was queued live and verified:
| Task | Result |
|---|---|
| Always-on Mac setup | ✅ Done — LaunchAgent + caffeinate configured |
| DigitalOcean hosting research | ✅ Done — s-2vcpu-4gb @ $24/mo documented in vault |
| Career search (SWE intern 2027) | ✅ Done — 4 jobs found, resumes tailored |
| IBM recruiter email draft | ✅ Done — full draft written to vault |
| UTR Spring-Ford tennis roster | ✅ Done — 77 players, UTR ratings extracted |
| Birthday video iMovie template | ⚠️ Partial — iMovie opened, blocked by Chrome/Teams at z-index 999 |
| DigitalOcean computer use plan | ✅ Done — plan documented in vault |
Key bugs found and fixed during the audit:
- Worker stall — workers were blocking on verification; fixed by moving verification async into
_pending_verifications - Gemini removal — Gemini CLI was causing intermittent failures; removed entirely, Claude-only with 30s retry
- Task log ordering — history was unordered; now sorted by
mtimewithHH:MMtimestamps - Career jobs disappearing — page reload wiped search results; fixed by persisting to
career_jobs_cache.json
Tech Stack
| Layer | Tech |
|---|---|
| Backend | Python 3.10+, FastAPI, Uvicorn, asyncio |
| Frontend | HTML5, Vanilla CSS/JS, WebSockets |
| LLM | Claude Code CLI (Anthropic) |
| Browser automation | Playwright |
| Desktop control | PyAutoGUI, screencapture, AppleScript |
| Notifications | Telegram Bot API (python-telegram-bot) |
| Storage | Obsidian vault (local Markdown files) |
Prerequisites
- macOS (required for computer use features)
- Python 3.10+
- Claude Code CLI — installed and authenticated
- A Telegram bot token and your chat ID
Installation
git clone https://github.com/Taran132g/agentic-os.git
cd agentic-os
pip install -r requirements.txt
For the career agent's auto-fill: PAIS drives the in-Chrome browser agent
Gemini (preferred — "Ask Gemini" button in Chrome's toolbar is
text-labeled, OCR-friendly, and runs on your Google account quota separate
from your Claude subscription) or Claude for Chrome (fallback) via
tools/computer.py. The form fills inside your real Chrome (your profile,
your cookies). See BROWSER_AGENT.md for the workflow.
Create a .env file in the project root:
TELEGRAM_BOT_TOKEN=your_bot_token_here
TELEGRAM_CHAT_ID=your_chat_id_here
GITHUB_TOKEN=your_github_token_here # optional
VAULT_PATH=/path/to/your/obsidian/vault # optional, has a default
DASHBOARD_PORT=8000 # optional
Usage
python main.py
- Dashboard:
http://localhost:8000 - Telegram: Send
/task research the best Python async patternsto your bot
To reload the LaunchAgent after config changes:
launchctl unload ~/Library/LaunchAgents/com.taran.pais.plist
launchctl load ~/Library/LaunchAgents/com.taran.pais.plist
Project Structure
agentic_os/
├── main.py # Entry point: FastAPI server + async workers + WebSocket
├── orchestrator.py # General task runner (Claude Code CLI + lessons injection)
├── career_workflow.py # Job search → Claude-for-Chrome fill → jobs cache
├── personal_tasks_workflow.py # Personal task planning and research + lessons injection
├── telegram_bot.py # Bot commands, routing buttons, approval callbacks
├── config.py # Env var loading
├── lessons.md # Append-only self-correction log (read by all agents)
├── deploy.sh # Deployment helper for DigitalOcean
├── tools/
│ ├── llm.py # Claude Code CLI runner with 30s usage-limit retry
│ ├── computer.py # Mac desktop control (screenshot, click, OCR, etc.)
│ ├── approval.py # Approval gate for sensitive actions
│ ├── logger.py # Vault task logger (YYYY-MM-DD HH-MM filenames)
│ ├── vault.py # Obsidian vault read/write helpers
│ ├── web.py # Web search and fetch utilities
│ ├── github_tools.py # GitHub API helpers
│ └── browser_fill.py # Drives Gemini (or Claude) in Chrome to fill job apps
├── dashboard/
│ ├── index.html # Main dashboard
│ ├── career.html # Career workflow UI (cached jobs, Gemini button removed)
│ └── personal.html # Personal tasks UI (self-correction feedback panel)
└── agents/ # Agent class definitions
Security
- Single-user: the Telegram bot rejects messages from any chat ID other than yours
- Approval gate blocks all external side-effect actions until you explicitly approve
- Local-first: all data stays on your machine; no data is sent to third-party services beyond the LLM APIs
License
MIT