Oh My Claudecode (OMC) is a multi-agent orchestration plugin for Claude Code. It adds 19 specialized agents for architecture, research, design, testing and data analysis, a set of skills and prompt triggers, and in-session modes that take a task described in natural language, split it across agents, run them in parallel where possible, and keep iterating until completion is verified. Model routing sends simple work to Haiku and complex reasoning to Opus, which the project reports as a 30 to 50% token saving, reusable problem-solving patterns are extracted from sessions, and a HUD statusline shows live orchestration metrics.
Team is the canonical orchestration surface and runs as a staged workflow: team-plan, team-prd, team-exec, team-verify, then a team-fix loop. It has two runtimes. The /team skill inside a Claude Code session uses Claude Code's native agent teams on a shared task list, and the omc team terminal command launches real CLI workers (Claude, Codex, Gemini, Antigravity, Grok or Cursor agent) in tmux panes that spawn on demand and exit when done. Other modes cover Autopilot (a single lead agent end to end), Execute (persistent verify and fix loops from plan to working code), Verify (repeated quality-gate cycling until tests, build, lint and typecheck pass), Ralph (persistence until verified completion), Ralplan (iterative planning consensus) and Deep Interview (Socratic questioning that measures clarity across weighted dimensions before any code is written). An artifact-only Ultragoal writes durable goal, checkpoint and evidence artifacts without starting a loop.
A provider advisor (omc ask in the terminal, /ask in a session) runs local provider CLIs for Claude, Codex, Gemini, Antigravity, Grok and Cursor and saves a markdown artifact, and a tri-model mode has Claude synthesize Codex and Antigravity output. The Codex and Gemini MCP servers were removed in v4.4.0 in favor of these CLI workers. Utilities include a pre-flight danger scan that reports risk findings before an unattended run, a daemon that resumes a session when a Claude rate limit resets, a stateful autoresearch loop with per-iteration evaluation and decision logs, session summaries and replay logs, stop callbacks to Telegram, Discord and Slack with tag lists, and forwarding of session events to an OpenClaw gateway.
The npm package also exports TypeScript helpers such as createOmcSession() for local Node.js programs built on the Claude Agent SDK, and the terminal commands are the path for CI and headless automation, since the slash skills need an active session. Runtime state, plans, logs and handoffs live in a per-repository state directory that several repositories can share as one workspace, and project-scoped skills in that directory are committable. A sibling project, Oh My Codex, provides the same orchestration for Codex CLI.
Features
- Team orchestration: a staged team-plan, team-prd, team-exec, team-verify and team-fix workflow on a shared task list, as an in-session skill or as terminal-launched tmux CLI workers
- 19 specialized agents: architecture, research, design, testing and data analysis agents with tier variants and a model-to-agent compatibility matrix with premium, balanced and budget presets
- Model routing: Haiku for simple tasks and Opus for complex reasoning, reported as a 30 to 50% token saving
- Execution modes: Autopilot, Execute, Verify, Ralph persistence, Ralplan planning consensus and Deep Interview requirements clarification
- Provider advisor: runs Claude, Codex, Gemini, Antigravity, Grok and Cursor CLIs for cross-checks and saves each answer as a markdown artifact
- Prompt triggers: words such as ralph, ralplan, deepsearch, ultrathink and stopomc switch modes without a slash command
- HUD statusline: real-time orchestration metrics in the status bar, plus session summaries, replay logs and token cost analytics
- Skill learning: extracts reusable patterns from sessions, with custom project skills kept in a committable directory
- Pre-flight danger scan: scans a task briefing and workspace before an unattended run and reports findings by severity
- Rate limit wait: a daemon that resumes Claude Code sessions when rate limits reset
- Notifications: stop callbacks to Telegram, Discord and Slack with tag lists, and session events forwarded to an OpenClaw gateway
- Library surface: TypeScript helpers such as createOmcSession() for programs built on the Claude Agent SDK
