Skip to content

iFrescoo/gart

BranchesTags

Repository files navigation

GART — Generative Agent Runtime Toolkit

Scaffold multi-IDE agentic coding projects with 144+ AI agents in one command.

npx gart-build my-project

GART sets up a complete AI agent orchestration environment for Claude Code, OpenCode (Google Gemini), and AntiGravity IDE — with pre-configured agents, skills, hooks, scoped rules, and MCP server integrations.

Quick Start

npx gart-build my-project
cd my-project
cp .env.example .env     # Add your API keys
claude                   # or: opencode / open in AntiGravity

The interactive CLI lets you:

  • Choose your tools — select which AI IDEs to configure (space to toggle)
  • Set agent language — agents respond in your preferred language
  • Auto-scaffold — only the configs you need, clean and ready

What You Get

Component Details
274 Specialist Agents Engineering, Design, Game Dev, Marketing, Sales, QA, and more
Agent Orchestration Team coordinators, division routing, central orchestrator
19 Skills Frontend design, debugging, brainstorming, git workflow, LSP navigation
23 Scoped Rules TypeScript, React, security, performance, accessibility, API design, and more
11 Hooks Dangerous command blocker, env leak protection, secret detection, auto-format, audit trail
MCP Pre-config GitHub, Memory, Playwright servers ready out of the box (.mcp.json)
Documentation Guides, ADR templates, task tracking, agent workflow reference

How GART Works

End User Workflow

flowchart LR
    A["npx gart-build my-project"] --> B["Select tools & options"]
    B --> C["Scaffold project"]
    C --> D["Follow SETUP.md"]
    D --> E{Selected tool}
    E --> F["claude — Claude Code"]
    E --> G["opencode — OpenCode"]
    E --> H["AntiGravity IDE"]
    F --> I["274+ agents ready to work"]
    G --> I
    H --> I
Loading

For end users — install, configure, work. No repo cloning needed.

Contributor Workflow

flowchart LR
    A["Fork & clone repo"] --> B["Edit agents, rules, skills"]
    B --> C["bash scripts/gart.sh sync"]
    C --> D["Test locally"]
    D --> E["Submit PR"]
    E --> F["CI validates"]
Loading

For contributors — fork, modify, sync, test, PR. Full access to docs-dev/, cli/, .github/.

Tool Switching Architecture

flowchart TB
    subgraph Shared["Shared (all tools)"]
        D["docs/"]
        S["scripts/gart.sh"]
        E[".env"]
    end
    subgraph CC["Claude Code"]
        CCA[".claude/ (agents, rules, hooks, skills)"]
        CCM["CLAUDE.md"]
    end
    subgraph OC["OpenCode"]
        OCA[".opencode/ (plugins, agents)"]
        OCM["AGENTS.md + opencode.json"]
    end
    subgraph AG["AntiGravity"]
        AGA[".agent/ (skills, workflows)"]
        AGM["GEMINI.md"]
    end
    Shared --- CC
    Shared --- OC
    Shared --- AG
Loading

Switch freely — shared docs and scripts, tool-specific configs isolated.

Supported Tools

Tool Config Files What's Included
Claude Code .claude/, CLAUDE.md, .mcp.json 274 agents, 19 skills, 11 hooks, 25 rules
OpenCode .opencode/, opencode.json, AGENTS.md 154 agents, 25 skills, commands
AntiGravity .agent/, GEMINI.md 155 skills, 25 rules, 6 workflows

Scoped Rules (Claude Code)

GART includes 23 per-domain rule files in .claude/rules/ — loaded automatically when Claude works with matching files:

Category Rules
Base core, code-generation, navigation, security, boundaries, mcp-tools, git, performance, error-handling, documentation
Languages typescript, python, css, html
Frameworks react, nextjs, tailwind
Testing testing, e2e
API & DB api-design, database
DevOps docker, ci-cd
Quality accessibility, code-review

Hooks (Claude Code)

11 hooks in .claude/hooks/ — automated safety, quality, and audit:

Hook Type What it does
block-dangerous-commands Security Blocks rm -rf, git push --force, DROP TABLE
prevent-env-leak Security Blocks cat .env, printenv, secret exposure
protect-files Security Blocks editing .env, credentials, .git/
detect-secrets Security Blocks hardcoded API keys, tokens, passwords
validate-commit-msg Quality Enforces Conventional Commits format
auto-format Quality Runs Prettier after edits
typecheck Quality Runs tsc --noEmit after TS edits
audit-trail Audit Logs every action to JSONL
bash-logger Audit Logs bash commands with timestamps
dep-audit Workflow Runs npm audit after package.json changes
warn-large-files Workflow Warns when files exceed 500 lines

Auto-Approve Permissions (Claude Code)

GART pre-configures .claude/settings.json with sensible auto-approve rules so Claude doesn't ask for confirmation on every file read, edit, or search. Dangerous commands (rm -rf, git push --force, git reset --hard) are explicitly denied.

To reduce prompts in your own projects, copy this to your global settings (~/.claude/settings.json):

{
  "permissions": {
    "allow": [
      "Read",
      "Edit",
      "Write",
      "Glob",
      "Grep",
      "WebSearch",
      "WebFetch",
      "Bash(git *)",
      "Bash(npm *)",
      "Bash(npx *)",
      "Bash(node *)",
      "Bash(ls *)",
      "Bash(mkdir *)",
      "Bash(cp *)",
      "Bash(mv *)",
      "Bash(bash *)",
      "Bash(echo *)",
      "Bash(find *)",
      "Bash(grep *)"
    ],
    "deny": [
      "Bash(rm -rf *)",
      "Bash(git push --force *)",
      "Bash(git reset --hard *)"
    ]
  }
}

Hooks still enforce safety (secret detection, env leak protection, dangerous command blocking) even with auto-approve enabled.

Scoped Rules & Workflows (AntiGravity)

GART includes the same 23 per-domain rule files in .agent/rules/ — loaded automatically when AntiGravity works with matching files. Same categories as Claude Code rules above.

6 workflows in .agent/workflows/ — triggered via / command:

Workflow What it does
orchestrate Multi-agent delegation across specialist domains
code-review Structured review: OK / Needs improvement / Critical
debugging 6-step protocol: reproduce, isolate, inspect, hypothesize, fix, prevent
git-workflow Branch naming, Conventional Commits, PR checklist
testing Test strategy: unit, integration, e2e with coverage targets
security-audit OWASP Top 10 checklist, dependency audit, secrets check

Prerequisites

Agent Divisions

Agents are organized into specialized divisions:

Division Examples
Engineering Backend Architect, Frontend Developer, DevOps, AI Engineer, Security
Design UX Architect, UI Designer, Brand Guardian, Visual Storyteller
Game Dev Unity, Unreal, Godot specialists, Level Designer, Narrative Designer
Marketing Content Creator, SEO, Social Media, Growth Hacker
QA & Testing API Tester, Performance Benchmarker, Evidence Collector
Specialized MCP Builder, Blockchain Auditor, Data Engineer, Technical Writer

Update agents to the latest version:

bash scripts/gart.sh sync

How It Works

  1. npx gart-build downloads the latest template from this repository
  2. You select which AI coding tools you use (Claude Code, OpenCode, AntiGravity)
  3. You choose your preferred agent response language
  4. GART scaffolds only the selected configs into your project
  5. Generated package.json, README.md, and .mcp.json are tailored to your selection

Contributing

This repository is the source for both the GART CLI and the template it scaffolds.

git clone https://github.com/iFrescoo/gart.git
cd gart

# CLI development
cd cli
npm install
npm run build
npm link              # Now 'gart-build' command is available globally

# Test scaffolding
gart-build test-project

Repository Structure

├── cli/              # GART CLI source (published to npm as "gart-build")
├── .claude/          # Claude Code agents, skills, hooks, rules
│   ├── agents/       # 274 specialist agents
│   ├── skills/       # 19 skills + orchestration
│   ├── hooks/        # 11 safety & quality hooks
│   └── rules/        # 23 scoped rules (per-domain)
├── .opencode/        # OpenCode agents, skills, commands
├── .agent/           # AntiGravity skills, workflows
├── scripts/          # Agent sync pipeline
├── docs/             # Documentation templates
├── CLAUDE.md         # Claude Code instruction file
├── AGENTS.md         # OpenCode instruction file
├── GEMINI.md         # AntiGravity instruction file
├── ROADMAP.md        # Future plans
└── marketplace.json  # Plugin marketplace metadata

Publishing

cd cli
npm run build
npm publish           # Requires npm login

Or push a version tag for automatic publishing:

git tag v1.1.0
git push origin v1.1.0   # GitHub Action publishes to npm

Acknowledgements

GART builds on top of the excellent work from:

  • msitarzewski/agency-agents — The upstream source for 274+ specialist AI agents, organized by division and role. GART's sync pipeline (gart.sh sync) pulls, enriches, and deploys these agents across Claude Code, OpenCode, and AntiGravity IDE.
  • Anthropic — Claude Code and the Claude API powering the agent runtime.
  • OpenCode — Multi-model AI coding assistant with plugin architecture.
  • mksglu/context-mode — Privacy-first MCP server for context window protection. Keeps raw tool output in a sandbox, reducing context consumption by ~98%.
  • vudovn/antigravity-kit — Community agent/skill/workflow kit for Google AntiGravity IDE. Inspired GART's .agent/ structure with specialized agents, skills, and workflows.
  • mixedbread-ai/mgrep — Semantic code search CLI by Mixedbread AI. Used in GART's OpenCode setup for intent-based code exploration alongside traditional grep.
  • skillsmp.com — Community skill marketplace with 280K+ indexed Agent Skills. Browse and discover skills for Claude Code, OpenCode, and more.
  • skills.sh — Vercel-backed skill leaderboard with install telemetry tracking across 19+ AI coding agents.
  • vava-nessa/free-coding-models — CLI that pings 174 free LLM models across 23 providers in real-time, ranks by stability and latency, and writes config for coding tools. Powers bash scripts/gart.sh models.

License

MIT

About

No description, website, or topics provided.

Resources

Readme

License

MIT license

Contributing

Contributing
Activity

Stars

2 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases 5

v1.3.0 Latest
Mar 22, 2026
+ 4 releases

Packages

 
 
 

Contributors

Languages