AI Personal OS
Atman36/ai-personal-os
Plano de controle para agentes de IA com estado, aprovações e tasks em JSON
Ver no GitHubFramework de sistema operacional pessoal para IA que fornece um plano de controle baseado em Markdown/JSON, agentes com roles, skills reutilizáveis e delegação com aprovação segura. Integra-se com Codex e Claude Code, mantendo estado explícito em vez de improvisar a partir do histórico de chat.
Instalação
npx convez-ai Atman36/ai-personal-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
- Orquestração de agentes IA com fila de execução controlada
- Gerenciamento de projetos e tarefas em Markdown/JSON com IA
- Delegação segura com aprovação e registro de autonomia
- Montagem de contexto e workflows de revisão semanal
Veredito

Framework robusto para quem quer montar um sistema de controle real para IA com estado durável, aprovações e trails de auditoria. Requer disciplina arquitetural, mas entrega um harness sólido além do mode chat.
README
A tradução para o português deste README ainda está na fila.
ai-personal-os
A Codex- and Claude Code-ready framework for a personal operating system: priorities, decisions, tasks, governed delegation, context hygiene, weekly review, and safe tool expansion.
What this framework is
This is not a prompt zoo and not a notes app replacement. It is a small control plane that lets an AI assistant operate against explicit state instead of improvising from chat history.
The framework combines:
- local Markdown/JSON as the v0 source of truth;
os/05_Control_Plane/active-work.jsonas the machine queue;- generated human mirrors such as
os/02_Planning/Task_Board.md; - role-based agents and reusable skills;
- local approval records, risk, autonomy, evidence-review, and self-improvement loops;
- a clear escalation path from local Codex/Claude Code to a durable harness such as Hermes.
The design direction comes from the research in research/: Personal OS should evolve from “chat over notes” into a controlled execution system with context assembly, memory, workflow state, approvals, telemetry, and replayable runs.
Install
cd ai-personal-os
python3 scripts/osctl.py init
python3 scripts/osctl.py validate-all
python3 scripts/osctl.py sync
Then open this folder in Codex (reads AGENTS.md) or Claude Code (reads CLAUDE.md). Both surfaces share the same skills through .agents/skills/ and .claude/skills.
First prompt
In Codex:
Открой $personal-os-gateway и проведи daily start.
In Claude Code: invoke the /personal-os-gateway skill or say «проведи daily start».
Useful prompts
Открой $personal-os-gateway и скажи, что мне делать дальше.
Открой $personal-os-decision-brief. Помоги решить, какой проект Personal OS должна вести первым.
Разложи эту инициативу в task contract, добавь в active-work.json и запусти python3 scripts/osctl.py sync.
Прочитай docs/HARNESS_DECISION.md и скажи, эту задачу запускать локально в Codex или в harness.
Key files
AGENTS.md— global instructions for Codex.CLAUDE.md— global instructions for Claude Code.docs/FRAMEWORK.md— framework architecture, invariants, and v0/v1 boundary.docs/HARNESS_DECISION.md— when local Codex/Claude Code is enough vs when Hermes-style harness is required.docs/LOCAL_AGENT_TASKS.md— implementation task-spec package for local agents.os/01_Operating_System/Naming_and_Context.md— naming convention and context hygiene rules..codex/agents/— Codex custom subagents..claude/agents/— the same role catalogue for Claude Code..agents/skills/— active skills shared by both surfaces.os/now.md— current focus.os/projects.md— active project truth.os/05_Control_Plane/active-work.json— machine-readable task queue.os/02_Planning/Task_Board.md— generated human board.scripts/osctl.py— init/validate/sync/status/add-task/decision/naming helper.research/— source analyses; prefer.mdfor Mermaid diagrams and.docxonly when source links are needed.
Core commands
python3 scripts/osctl.py init
python3 scripts/osctl.py status
python3 scripts/osctl.py validate
python3 scripts/osctl.py validate-all
python3 scripts/osctl.py sync
python3 scripts/osctl.py note-name --type research --description "MCP server comparison" --date 2026-03-18
python3 scripts/osctl.py add-task --title "Example" --next-step "Do the first slice" --done-when "Result is reviewable"
python3 scripts/osctl.py decision --title "Choose first managed project" --option "AI-MAX growth" --option "Personal productivity" --option "Agent corporation"
python3 scripts/osctl.py approval-request --task-id task-1 --requested-by ai_ops_lead --requested-for github_issue_create --policy-action pause_for_founder_approval --tool-scope github:issues --summary "Create a GitHub issue for task-1" --risk-tier medium --autonomy-tier A3
python3 scripts/osctl.py approval-list
python3 scripts/osctl.py approval-decide --approval approval-id --decision approved --decided-by user
Local Codex vs harness
Use local Codex/Claude Code for small, reversible A0-A2 work: docs, repo-local edits, tests, decision briefs, task contracts, and state sync.
Use a harness such as Hermes when the work needs any of these: durable checkpoints, replay, persistent workspace, parallel agents, external connectors, side effects, approval records, trace/artifact bundles, long-running execution, or recovery after interruption.
The rule of thumb: if losing the chat would lose the work, or if the task can affect the outside world, run it through a harness.
Privacy
This repo is meant to be public, so it ships the framework only. Your personal content is kept out of git:
- Live state (
os/now.md,os/projects.md,os/05_Control_Plane/active-work.json, planning and decision files) is gitignored and seeded from committed*.exampletemplates. - Private runtime notes, handoffs, approvals, artifacts, and telemetry live in
.os_runtime/(gitignored). - Build/meta notes about creating the OS itself (
SESSION_*.md,/_meta/,/private/, etc.) are gitignored.
See .gitignore for the full list. Run python3 scripts/osctl.py init after cloning.
Operating rule
Do not try to make the OS autonomous on day one. Make it useful first: daily start, decision briefs, task contracts, weekly review, context hygiene, and controlled delegation. Add connectors only after the local control plane is coherent and validated.