agentic-os

TL;DR: An agentic personal operating system built to automate high-leverage workflows across Claude Code, Codex, Pi, OpenClaw, and other coding agents/runtime platforms.

Quick Start

1. Clone this repo

   git clone https://github.com/itseffi/personal-os.git
   cd personal-os
   

2. Run setup

   chmod +x setup.sh
   ./setup.sh
   

3. Start using This automates high-leverage execution end-to-end: it converts raw backlog into prioritized, goal-aligned, verification-enforced action plans.

   Open this repo in your agent and run:
   1) "Process my backlog from BACKLOG.md into Tasks/**/*.md using AGENTS.md rules."
   2) "Show my P0/P1 unblocked tasks aligned to GOALS.md."
   3) "Propose today’s top 3 with required verification evidence and commands."
   

Quick Links

Build Your Personal OS · Workflows · Canonical Skills · Evals · Tutorials (index)

Architecture

flowchart TD
    U["User Prompt"] --> A["Agent Runtime<br/>Claude Code | Codex | Pi | OpenClaw"]
    A --> I["Instructions<br/>AGENTS.md + wrappers"]
    A --> S["Skills<br/>.agents/skills/*/SKILL.md"]
    A --> W["Workflows<br/>Workflows/*.md"]
    A --> F["State + Context<br/>Tasks, GOALS, BACKLOG, Knowledge, Resources"]
    A -. optional .-> M["MCP Integrations<br/>System/mcp + external services"]
    A -. optional .-> D["Subagents<br/>runtime-dependent delegation"]
    A --> E["Evals<br/>Evals/ + Evals/skills + scripts/run_skill_evals.py"]

    classDef core fill:#ff9891,stroke:#2b2b2b,color:#111111,stroke-width:1.2px;
    classDef optional fill:#ffd4d0,stroke:#2b2b2b,color:#111111,stroke-width:1.2px,stroke-dasharray: 4 3;
    class U,A,I,S,W,F,E core;
    class M,D optional;

Agent Compatibility

Personal OS is designed to work with Claude Code, Codex, Pi, OpenClaw, and similar coding agent runtimes.

• Shared behavior: AGENTS.md
• Claude wrapper: CLAUDE.md
• Codex wrapper: CODEX.md
• Pi wrapper: PI.md
• OpenClaw wrapper: OPENCLAW.md
• Canonical runtime skills: .agents/skills/*/SKILL.md
• Skills in this repo follow the Agent Skills open standard.
• This repo uses skills with progressive disclosure to manage context efficiently: agents begin with each skill's metadata (name, description, file path, plus agents/openai.yaml), and load full SKILL.md instructions only when a skill is selected. Canonical skills live in .agents/skills/, with bridge paths for Claude, Pi, and OpenClaw.
• Optional subagents are supported when the runtime provides agent delegation features (not required for core repo operation).
• Claude bridge path: .claude/skills -> ../.agents/skills (symlink)
• Pi bridge: configure Pi to point to this repo and use .agents/skills/ as its skill source
• OpenClaw bridge: create skills -> .agents/skills symlink (or load .agents/skills via OpenClaw config)

Bridge bootstrap (run once from repo root):

mkdir -p .claude
ln -sfn ../.agents/skills .claude/skills
ln -sfn .agents/skills skills

For Codex/OpenAI-style routing metadata, this repo includes:

• .agents/skills/<skill>/agents/openai.yaml (Claude, Pi, and OpenClaw primarily use SKILL.md and do not require this file format.)

Pi Local/Offline Setup (Optional)

You can run Personal OS with Pi using a local/offline model backend (for example llama.cpp) or a hosted endpoint. For full setup instructions (server launch, ~/.pi/agent/models.json, and runtime configuration), see Pi Agent Setup.

File System Layout

personal-os/
├── AGENTS.md           # AI agent instructions (the brain)
├── GOALS.md            # Your goals and priorities
├── BACKLOG.md          # Quick capture inbox
├── Tasks/              # Your active work
├── Knowledge/          # Your notes and docs
├── Resources/          # Voice samples, templates, references
├── Workflows/          # Daily + Product & Strategy workflows
├── .agents/skills/     # Canonical Codex/OpenAI skill packs
├── Evals/              # Session reviews
├── Tutorials/          # Learning guides
└── System/             # MCP server, templates, integrations

Semantics by location: Tasks/**/*.md = actionable work, Knowledge/**/*.md = reference context.

How It Works

The Memory Stack

AGENTS.md        →    Instructions layer (how AI behaves)
GOALS.md         →    Priority layer (what matters)
Tasks/**/*.md     →    State layer (current work)
Knowledge/**/*.md →    Context layer (reference)
.agents/skills/* →    Capability layer (how the agent executes specialized workflows)

Privacy First

Personal operating data stays local (gitignored):

• Tasks/ - your work
• Knowledge/ - your notes
• Resources/ - your samples
• BACKLOG.md - your inbox

Some top-level configuration files (AGENTS.md, GOALS.md, CLAUDE.md, CODEX.md, PI.md, OPENCLAW.md, docs) are version controlled by design. Treat GOALS.md as potentially sensitive and review content before publishing a public repo.

Evals

This repo includes structural, behavioral, routing, and memory-impact evals.

Run:

python scripts/validate_skills.py
python scripts/validate_skill_eval_cases.py
python scripts/run_skill_evals.py --provider fixture
python scripts/run_routing_evals.py
python scripts/run_memory_impact_evals.py

Optional live-model run (OpenAI-compatible endpoint, local or remote):

python scripts/run_skill_evals.py --provider openai --model your-model-id

Outputs are written to:

• Evals/skills/results/
• Evals/memory/results/

Use these evals as a regression gate when updating .agents/skills/.

Long-Running Agent Principles

Personal OS follows four operating patterns:

• Skills: versioned procedures in .agents/skills/*/SKILL.md
• Shell execution: run real tasks in terminal environments and produce artifacts
• Compaction-aware workflows: structure long runs to preserve continuity
• Verification-first completion: require fresh evidence before claiming work is done

Security defaults:

• Keep network access minimal and allowlist-based
• Treat tool output as untrusted input
• Use explicit review boundaries for generated artifacts

Tech Stack

• File Format: Markdown with YAML frontmatter
• Agent Runtimes: Claude Code, Codex, Pi, OpenClaw, Cursor, and similar coding agent runtimes
• Optional Integrations: MCP (Slack, Linear, Google Calendar, Atlassian, Granola)
• Version Control: Git

Contributing

Issues and PRs welcome.