A weapon, not a tool
Multi-AI orchestration plugin for Claude Code. Coordinate Claude, Gemini, and Codex with 19 specialized agents, 47 skills, and 55 MCP-powered tools.
Two new opt-in shipyard skills, remote approval gates with workspace checkpoints and rollback for graph, HUD and Windows cache performance work, and hook protocol fixes.
graphThe graph workflow gains remote approval gates plus workspace checkpoints and rollback, so a staged change can be paused for sign-off and reverted to a known-good workspace state.
loftShipyard shape-before-steel discipline: shape the work before committing to implementation. Opt-in and model-invoked, so it only engages when the model judges it useful.
ask-navigatorA shipyard navigator for foggy efforts — it takes an unclear objective and drives it toward a concrete next step. Opt-in.
Git path memoization is now scoped per render, so HUD refreshes stop repeating filesystem work across unrelated frames.
Hook stderr survives an early protocol stdout close, and the unsupported suppressOutput field is omitted from PostToolUse payloads.
Windows cache occupancy identity checks are batched and occupancy tick precision is preserved, cutting redundant probes without losing accuracy.
Claude, Codex, and Gemini CLI workers running in parallel tmux panes — each doing what it does best.
The Conductor
Spawns a real gemini CLI process in a tmux pane. 1M token context for design review and UI work across your entire codebase.
Spawns a real codex CLI process in a tmux pane. Deep code analysis, architecture validation, and security review.
Claude decomposes tasks, fans out work to parallel worker lanes via the team skill (omc team for legacy tmux/CLI runs), and synthesizes results — no MCP servers required.
Install via Claude Code plugin marketplace, or use the npm package directly: [email protected].
Install the plugin
# In Claude Code, run:
/plugin marketplace add https://github.com/Yeachan-Heo/oh-my-claudecode
/plugin install oh-my-claudecode
Or install the npm package directly
# CLI install via npm:
npm install -g [email protected]
Run setup
/oh-my-claudecode:omc-setup
# That's it! You're ready.
Choose how you want to work. From full autopilot to token-efficient parallel execution.
Full autonomous execution from high-level idea to working, tested code. No manual intervention required.
Self-referential loop until completion with architect verification. The boulder never stops rolling.
Carries an approved task through to working, verified code, with team owning the parallel worker lanes.
Socratic questioning to clarify vague ideas before execution. Exposes hidden assumptions and measures clarity.
N coordinated agents on a shared task list using Claude Code native teams. Real-time messaging and task coordination.
Strategic interview-based planning workflow. Gather requirements before diving into implementation.
Language server integration, structural code analysis, and persistent state - all built in.
Hover info, go-to-definition, find references, and project-wide type checking with any language server.
Structural code search and replace. Find patterns by syntax, not just text matching.
Sandboxed execution: imports, file I/O, and dynamic code execution are blocked; use built-in functions and persistent variables for calculation and in-memory data analysis.
Notepad, project memory, and session state that persists across conversations.
Four keywords activate automatically. Everything else is an explicit invocation — team in particular has no trigger, and the old plan the … activation was removed.
| Keyword | Effect | Example |
|---|---|---|
autopilot keyword |
Full autonomous execution | "autopilot: build a todo app" |
ralph keyword |
Persistence mode until completion | "ralph: refactor auth system" |
/oh-my-claudecode:ultragoal explicit |
Durable goal ledger that runs to completion | "ultragoal: ship the billing rewrite" |
/oh-my-claudecode:execute explicit |
Maximum parallelism across worker lanes | "execute: fix all type errors" |
/oh-my-claudecode:autoresearch explicit |
Goal-directed research mission ending in a verdict | "autoresearch: which queue fits our load" |
/oh-my-claudecode:graph explicit |
Remote approval gates with workspace checkpoints and rollback | "graph: stage the migration" |
/oh-my-claudecode:omc-plan explicit |
Planning interview mode | "/oh-my-claudecode:omc-plan the new API" |
ralplan keyword |
Iterative planning consensus | "ralplan this feature" |
deep-interview keyword |
Socratic requirements clarification | "/deep-interview 'vague idea'" |
/oh-my-claudecode:team explicit |
Coordinated native teams (replaces swarm/ultrapilot) | "team 5:executor fix all errors" |
swarm Deprecated |
Use team instead |
"team 5:executor fix all errors" |
ulw / ultrawork Retired |
Retired mode; execute handles parallel lanes |
"execute: fix all type errors" |
ultrapilot Retired |
Routes to execute |
"execute: land the refactor" |
Unified agents organized by domain. Use the model parameter to match complexity.
Core workflow from exploration to verification
Multi-dimensional code quality checks
Deep expertise for specialized tasks
All agents use model-optimized routing (Opus for architecture/analysis, Sonnet for standard work, Haiku for quick lookups).
No docs needed. No commands to memorize. Just say what you want.
"Build me a REST API" - that's it. Autopilot detects your intent and orchestrates 19 agents automatically.
Say "fast" and it activates parallelism. Say "don't stop" and it persists. Every intent maps to the right mode.
Delegates to specialists, routes to the right model tier, verifies with architect. You focus on the what, not the how.
Install from Claude's plugin marketplace, run /oh-my-claudecode:omc-setup - done. Everything else is automatic.