oh-my-claudecode (Yeachan-Heo) — Summary

oh-my-claudecode by Yeachan-Heo is the flagship "oh-my-*" framework with 34,919 stars — by far the most-adopted in this batch and one of the highest-starred AI coding frameworks in existence. It ships as an npm package (oh-my-claude-sisyphus v4.14.4) with three CLI binaries (oh-my-claudecode, omc, omc-cli), 19 named agents (architect, executor, planner, critic, analyst, researcher, scientist, designer, writer, qa-tester, security-reviewer, code-reviewer, explore, vision, verifier, tracer, debugger, document-specialist, code-simplifier), 28+ skills (autopilot, ralph, ralplan, ultrawork, ultraqa, deep-interview, team, plan, hud, omc-setup, and more), and a JavaScript hooks system with 8+ hook modules. The signature features: Team mode (v4.1.7+) spawns tmux CLI workers for Claude/Codex/Gemini in parallel; autopilot runs a full 6-phase lifecycle from idea to verified code; ralph implements a PRD-driven persistence loop; deep-interview uses Socratic questioning with mathematical ambiguity scoring; ralplan implements a 3-stage consensus plan (planner→architect→critic). The project has an active Discord community and dedicated documentation site.

Differs from seeds: Closest to claude-flow (npm package + CLI + large agent roster + complex orchestration) but exceeds it on every dimension: 19 agents (vs 44 but claude-flow uses code-class, OMC uses persona-md), 28+ skills (vs 107), a Team mode with tmux multi-CLI spawning not present in any seed, PRD-driven ralph loops with persistence, and Socratic deep-interview with mathematical ambiguity scoring — all absent from the 11 seeds.

oh-my-claudecode (Yeachan-Heo) — Overview

Origin

Created by Yeachan Heo, inspired by oh-my-opencode. Published on npm as oh-my-claude-sisyphus (note: different from the GitHub repo name for historical reasons). At v4.14.4 with 34,919 stars, 3,192 forks, and an active Discord (5,000+ members implied by server).

"Multi-agent orchestration for Claude Code. Zero learning curve."

"Don't learn Claude Code. Just use OMC."

Philosophy

Teams-first multi-agent orchestration:

"Autopilot: build a REST API for managing tasks — That's it. Everything else is automatic."

"If teams are disabled, OMC will warn you and fall back to non-team execution where possible."

The system prioritizes zero learning curve — users invoke /autopilot "build X" and the framework handles the entire lifecycle.

Key Manifesto Quotes

"For Codex users: Check out oh-my-codex — the same orchestration experience for OpenAI Codex CLI."

"Teams is the canonical orchestration surface in OMC (v4.1.7+). The legacy swarm keyword/skill has been removed."

"Team runs as a staged pipeline: team-plan → team-prd → team-exec → team-verify → team-fix (loop)"

The "sisyphus" Naming

The npm package is named oh-my-claude-sisyphus while the repo is oh-my-claudecode. Per README: "The project is branded as oh-my-claudecode (repo, plugin, commands), but the npm package is published as oh-my-claude-sisyphus."

Community

• Discord: https://discord.gg/PUwSMR9XNk
• Documentation site: https://yeachan-heo.github.io/oh-my-claudecode-website
• Multi-language README: English, Korean, Chinese, Japanese, Spanish, Vietnamese, Portuguese

oh-my-claudecode (Yeachan-Heo) — Architecture

Distribution

• Type: npm package + Claude Code plugin
• npm package: oh-my-claude-sisyphus
• Version: 4.14.4
• Stars: 34,919 (2026-05-26)

Install

# Via Claude Code marketplace (recommended)
/plugin marketplace add https://github.com/Yeachan-Heo/oh-my-claudecode
/plugin install oh-my-claudecode

# Setup
/setup   # or /omc-setup inside Claude Code session
omc setup  # from terminal

# Via npm (CLI path)
npm i -g oh-my-claude-sisyphus@latest

CLI Binaries

"bin": {
  "oh-my-claudecode": "bridge/cli.cjs",
  "omc": "bridge/cli.cjs",
  "omc-cli": "bridge/cli.cjs"
}

Directory Tree (high-level)

oh-my-claudecode/
├── agents/           # 19+ agent markdown files
├── skills/           # 28+ skill directories
├── commands/         # Slash command definitions
├── hooks/
│   └── hooks.json    # Hook declarations (JS/Node)
├── bridge/
│   ├── cli.cjs       # CLI entry
│   ├── mcp-server.cjs  # MCP server
│   ├── team-bridge.cjs # tmux team coordination
│   ├── team-mcp.cjs
│   └── gyoshu_bridge.py
├── src/
│   ├── agents/       # TypeScript agent definitions
│   ├── hooks/        # 8 hook modules
│   ├── features/     # 6 feature modules
│   │   ├── model-routing/
│   │   ├── boulder-state/
│   │   ├── context-injector/
│   │   ├── background-agent/
│   │   └── builtin-skills/
│   └── installer/
├── templates/
└── .claude-plugin/

Key State Files

• .omc/state/sessions/{sessionId}/prd.json — PRD state (ralph loop)
• .omc/plans/ralplan-*.md — Ralplan consensus plans
• .omc/specs/deep-interview-*.md — Deep interview specs
• .omc/notepads/ — Agent learnings across iterations
• .sisyphus/ (legacy) — Legacy boulder state

Required Runtime

• Node.js ≥ 18
• better-sqlite3 (native addon, for sqlite state)
• tmux (optional, for Team mode CLI workers)
• codex / gemini CLIs (optional, for multi-CLI Team mode)

oh-my-claudecode (Yeachan-Heo) — Components

Agents (19+)

Multiple agents have low/medium/high suffix variants for model tier selection.

Skills (28+)

Hooks (JavaScript, 8+ modules)

Events: UserPromptSubmit, SessionStart, PreToolUse, PermissionRequest

Key hook scripts:

• keyword-detector.mjs — detects ultrawork/ralph/autopilot keywords
• skill-injector.mjs — injects skill context at prompt submit
• session-start.mjs — initializes session
• project-memory-session.mjs — loads project memory
• pre-tool-enforcer.mjs — pre-tool enforcement rules
• wiki-session-start.mjs — loads wiki on session start
• setup-init.mjs, setup-maintenance.mjs — setup hooks

CLI Commands

oh-my-claudecode (Yeachan-Heo) — Uniqueness

Differs from Seeds

Closest to claude-flow (npm package + CLI binary + TypeScript + 44 subagents + MCP) but exceeds it on multiple dimensions: (1) Team mode with tmux CLI workers — spawning Codex/Gemini/Claude CLI processes in tmux panes for heterogeneous multi-CLI orchestration is absent from all 11 seeds; (2) Ralplan 3-stage consensus (planner → architect → critic sign-off) is the closest thing to a formal consensus mechanism in the corpus outside of claude-flow's hive-mind protocols; (3) Deep interview with mathematical ambiguity scoring — Socratic ONE-question-per-round with quantified ambiguity gating before execution is unique; (4) Autopilot's state-skip logic (check for pre-existing ralplan/deep-interview specs before running phases) enables pipeline resumption without manual coordination; (5) the companyContext.tool MCP hook for enterprise customization is absent from all seeds.

The Star Gap

At 34,919 stars, this framework is 63x larger than the next in this batch (opensoft at 544). The gap reflects: (1) early-mover advantage as a Claude Code port of oh-my-opencode; (2) zero-learning-curve UX (/autopilot "build X"); (3) active Discord community; (4) regular version updates (v4.14.4 with active development).

Positioning

• The de facto standard for multi-agent Claude Code orchestration
• "Don't learn Claude Code. Just use OMC."
• Companion project: oh-my-codex for the same UX on Codex CLI

Observable Failure Modes

1. npm package naming confusion: Project is oh-my-claudecode but npm package is oh-my-claude-sisyphus — update commands use the wrong name for users who don't read carefully
2. better-sqlite3 native addon: Requires native compilation; prebuild-install deprecation warning already logged; CI breaks on ARM/musl targets
3. Team mode tmux dependency: omc team requires an active tmux session + codex/gemini CLIs installed — significant setup requirement not present in other frameworks
4. CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS: Native team mode requires experimental flag; if it changes behavior or is removed, /team breaks silently
5. Ralph PRD state drift: If a session ID changes between ralph iterations (session restart), PRD state at .omc/state/sessions/{oldId}/prd.json is orphaned with no automatic migration
6. Deep interview ambiguity threshold: Mathematical scoring is agent-evaluated, not programmatic — if the model underscores ambiguity, the gate is bypassed