SkillOS is a Meta Skill Operating System for AI agents. It helps agents manage other skills: list, route, rate, analyze, optimize, generate, compare, detect conflicts, map relationships, and recommend skill workflows.

SkillOS is intentionally not a replacement for domain skills. It does not write your React app, design your database, or deploy production infrastructure. It helps your agent choose, evaluate, create, and compose the right skills for those jobs.

As agents gain more skills, the hard problem becomes skill operations:

• Which skill should handle this request?
• Are two skills fighting over the same trigger words?
• Is a skill well written enough to auto-trigger reliably?
• What order should several skills run in?
• How do I install the same skill manager into Codex, Claude Code, or another agent without hand-copying random files?

SkillOS provides a deterministic CLI plus lightweight Agent adapters so every user can deploy the same capability without guessing what to copy.

SkillOS has two distinct surfaces:

1. Core CLI: the full Python project in this repository.
2. Agent adapter: a small SKILL.md package installed into an agent's skill directory. The adapter teaches the agent how to call the Core CLI.

Do not copy the whole repository into an agent skill folder. Use install so SkillOS installs only the files that each agent needs.

git clone https://github.com/K-arch-droid/SkillOS.git
cd SkillOS
python skillos.py doctor

Install SkillOS into an agent:

# Codex
python skillos.py install --agent codex

# Claude Code
python skillos.py install --agent claude-code

# Generic agent / custom target
python skillos.py install --agent generic --target /path/to/agent/skills/skillos

Verify the adapter:

python skillos.py doctor --agent codex

Preview without writing files:

python skillos.py install --agent codex --dry-run

Each adapter is generated from adapters/<agent>/ and bundled with the shared references/ and templates/ files it needs.

An Agent adapter contains:

skillos/
├── SKILL.md
├── agents/openai.yaml          # Codex only
├── references/
│   ├── REVIEW-CHECKLIST.md
│   ├── ROUTING-RULES.md
│   ├── SKILL-TYPES.md
│   ├── GOTCHAS.md
│   └── ADVANCED-PATTERNS.md
└── templates/
    └── EVAL-TEMPLATE.md

The adapter does not copy the full source tree, test suite, .git, runtime state, generated registries, or development artifacts. It points back to the Core CLI path resolved during installation.

SkillOS rates skills with seven weighted dimensions:

Grades: A (90-100), B (80-89), C (70-79), D (60-69), F (<60).

SkillOS/
├── skillos.py                  # CLI entry point
├── skill_parser.py             # SKILL.md parser
├── skill_analyzer.py           # 7-dimension rating engine
├── skill_router.py             # Request-to-skill router
├── skill_registry.py           # Installed skill scanner and registry
├── skill_generator.py          # Skill generator and templates
├── skill_optimizer.py          # Review-to-plan optimizer
├── conflict_detector.py        # Conflict and relationship intelligence
├── adapters/                   # Agent-specific install surfaces
│   ├── codex/
│   ├── claude-code/
│   └── generic/
├── references/                 # Shared rating/routing/design references
├── templates/                  # Skill and eval templates
├── tests/                      # Unit and CLI tests
└── SKILL.md                    # Project-level skill instructions

Run the full test suite:

python -m tests.test_all

Package an adapter:

python skillos.py package --agent codex -o dist/skillos-codex.zip

Check a local adapter install in a temporary directory:

python skillos.py install --agent codex --target ./tmp/skillos
python skillos.py doctor --agent codex --target ./tmp/skillos

• Adapter, not migration: each agent receives a lightweight adapter, not the whole repository.
• CLI is canonical: deterministic operations happen in Python, not in prompt text.
• Progressive disclosure: SKILL.md stays small; references load only when needed.
• Composition over fusion: SkillOS composes skills instead of merging them into mega-skills.
• Safe optimization: optimize plans first and only modifies after explicit confirmation.

MIT