Skip to content

About

Creator operating system for AI-native workflows: skills, commands, agents, and plugins for Codex, Claude Code, Cursor, Grok and Antigravity Gemini.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

12 stars

Watchers

1 watching

Forks

Latest commit

 

History

135 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Agentic Creator OS

The skills, agents and commands behind GenCreator, installable in Claude Code, Codex, Cursor, Grok and Antigravity.

Reusable skills, commands, agent profiles, and local safety tooling for AI-assisted creative work.

Quick start · Project instructions · Contributing · Issues

What this repository is

Agentic Creator OS (ACOS) is a public source repository for creator workflows that can be loaded into Claude Code or rendered as context for other coding agents. It contains:

  • domain skills written as SKILL.md modules;
  • top-level slash-command definitions for Claude Code;
  • specialist agent profiles;
  • local hook source and configuration;
  • generators for Cursor, Windsurf, Gemini, Grok, Antigravity, and generic context files;
  • MCP server workspaces and an observatory catalog.

ACOS is a source distribution, not a hosted service. The current version metadata is 11.0.0.

Verified public surface

The repository currently contains:

Surface Count Definition
Non-empty skill modules 175 Non-empty .claude/skills/**/SKILL.md files
Empty skill placeholders 5 Tracked SKILL.md paths with no content; not counted as modules
Top-level slash commands 85 .claude/commands/*.md, excluding the directory context file
Top-level agent profiles 81 .claude/agents/*.{md,json}, excluding the directory context file
Installable shell hooks 9 Top-level .claude/hooks/*.sh files copied by the Claude installer
Activation rules 32 Entries in .claude/skill-rules.json

These numbers are generated from the tree. npm run verify:public-surface fails if the counts, version metadata, install smoke test, or public license and marketplace statements drift.

Install from source

The verified installation path is the repository installer:

git clone https://github.com/frankxai/agentic-creator-os.git
cd agentic-creator-os
./install.sh --dry-run --platform=claude
./install.sh --platform=claude

--dry-run prints the files that would be created or replaced and writes nothing. The second command writes.

The Claude path copies the repository's skill groups, top-level commands, agent profiles, shell hooks, hook configuration, activation rules, and state metadata into CLAUDE_HOME (or ~/.claude by default).

Before installing into an existing profile, inspect the script and back up your profile. The Claude installer checks every file it would copy before writing anything: a differing file or symlinked destination stops the install instead of replacing your work. An unchanged repeat install skips identical files. It still writes ACOS-owned acos/state.json; there is no automatic uninstall or rollback yet. --target controls project files for other platforms, not the Claude profile path; set CLAUDE_HOME to choose an isolated Claude destination.

For an isolated inspection:

export CLAUDE_HOME="$(mktemp -d)"
./install.sh --platform=claude --minimal
find "$CLAUDE_HOME" -type f | sort

Other platform flags generate context or harness-specific files:

./install.sh --platform=cursor
./install.sh --platform=windsurf
./install.sh --platform=gemini
./install.sh --platform=grok
./install.sh --platform=antigravity
./install.sh --platform=generic

Those generators are present in the tree. They should not be read as a claim that every external harness or third-party connector has been independently certified end to end.

One plugin, five practices

The full source install above copies the whole catalog. For a first install, add this repository as a marketplace and install creator-os-core. It contains five skills and the /acos command: start safely, review a diff, write in a stated voice, run the project's checks, and hand off without secrets.

# Claude Code
claude plugin marketplace add frankxai/agentic-creator-os
claude plugin install creator-os-core@frankx-creator

# Codex
codex plugin marketplace add frankxai/agentic-creator-os
codex plugin add creator-os-core@frankx-creator

# Grok
grok plugin marketplace add frankxai/agentic-creator-os
grok plugin install creator-os-core --trust

Open a new session after installing. Type /acos, or say what you are doing in one sentence.

You have five practices: start safely, review a diff, write in a stated voice, run the checks, and hand off clean. The first things people ask are "Review this diff", "Write the install note", and "Run the checks".

If the new session does not know /acos, the marketplace was added in a different profile. Run the add command again in that profile, then install the plugin, then open another session.

The index is .claude-plugin/marketplace.json. Grok also reads that Claude-compatible index.

The separate agentic-creator-skills repository is a companion catalog. Evaluate its installation instructions and license independently.

Practice receipt

tools/receipt-check/check-receipt.mjs answers yes or no for one practice receipt. Yes means the five fields match the public rules. The checker sets signed to false. A yes is not a human signature and not a lab attestation.

node --test tools/receipt-check/check-receipt.test.mjs
node tools/receipt-check/serve.mjs

POST /check takes the receipt JSON. POST /mcp accepts initialize, tools/list, and tools/call for check_practice_receipt. The server listens on 127.0.0.1 and RECEIPT_PORT (default 8787). It is a local check, not a public registry.

How the pieces fit

flowchart LR
  Intent["Creator intent"] --> Instructions["Project instructions"]
  Instructions --> Skills["Skills"]
  Instructions --> Commands["Commands"]
  Instructions --> Agents["Agent profiles"]
  Skills --> Work["Work product"]
  Commands --> Work
  Agents --> Work
  Hooks["Local hook source"] --> Evidence["Local receipts and checks"]
  Work --> Evidence
Loading

The files provide instructions and tooling. Whether a skill activates, a command is available, or a hook executes depends on the target harness and its local configuration.

Start with the system

After installation, inspect these entry points:

CLAUDE.md                    Project behavior and command map
AGENTS.md                    Repository-specific agent instructions
.claude/skill-rules.json     Activation rules
.claude/commands/acos.md     Main router command
.claude/skills/              Skill source
.claude/agents/              Agent profiles
.claude/hooks/               Hook source

Useful first prompts:

Read CLAUDE.md and AGENTS.md, then show the ACOS capabilities relevant to this repository.
Plan a creator workflow using only skills that exist in .claude/skills.
Review the proposed workflow and identify every external dependency before execution.

Agent OS: your own C-suite of agents

Nine Generals (CEO, CTO, CMO, CFO, COO, CPO, CAIO, CHRO, CCO), compiled from one shared Expertise Kernel. Each keeps its own memory, delegates to skills and worker agents, reads your identity files, and stops at your gates. Optional Domain Queens route each brand's work to its Generals. A scanner scores every skill, agent, and command you can load, and a daily CI check keeps the whole setup from drifting. Dependency-free Node; no install step.

node scripts/agent-os-init.mjs --name <you>          # your instance from the template
node scripts/agent-compile.mjs --instance <you>      # your Generals
node scripts/estate-audit.mjs --repos <dir> --md audit.md

Start with docs/agent-os/ADOPT.md; the full design is in docs/agent-os/.

Verification

npm install --no-audit --no-fund
npm run verify:public-surface
npm run lint
npm run observatory:test
npm run build:all

The public-surface verifier performs an isolated Claude install smoke test. The main CI installs dependencies, verifies the public surface, builds and lints the MCP workspaces, and runs their type checks as a blocking gate.

Release discipline

Changes to public claims should answer four questions:

  1. What source file proves the claim?
  2. What check prevents it from drifting?
  3. What external dependency remains unverified?
  4. What named-owner decision is still required?

Large resets should be split by concern. Documentation truth, installer behavior, hook wiring, third-party code provenance, legal terms, and marketplace distribution have different review and approval requirements.

Related public systems

License status

Original software and technical documentation use Apache-2.0, unless a file or directory states different terms. See LICENSE, LICENSING.md, and NOTICE. Historical MIT grants and third-party licenses remain in effect; the license does not grant trademark rights.

Contributing

Read CONTRIBUTING.md and the nearest AGENTS.md or CLAUDE.md before editing. Keep changes narrow, preserve existing safety boundaries, and attach commands or artifacts that let another reviewer reproduce the result.

About

Creator operating system for AI-native workflows: skills, commands, agents, and plugins for Codex, Claude Code, Cursor, Grok and Antigravity Gemini.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

12 stars

Watchers

1 watching

Forks

Releases

Packages

Used by

Contributors

Languages