Multi-agent orchestration framework for Claude Code: 19 specialized agents, intelligent Haiku/Opus model routing, and an autonomous plan-to-ship pipeline with 30-50% token cost reduction.
Running every Claude Code task through the same general session means paying Opus prices for tasks Haiku handles fine, losing context on long builds, and manually managing multi-step workflows. OMC installs 19 specialized agents with intelligent model routing. Routing simple tasks to Haiku and complex architectural work to Opus. Its plan→design→execute→verify→fix pipeline runs autonomously with persistent loops and auto-retry. Autopilot mode handles well-defined tasks end-to-end without staged coordination. The result is 30-50% token cost reduction plus consistent quality gates on every phase.
Steps
Install via the Claude plugin marketplace. After install, restart Claude Code and check the version. OMC v4.7.7 removed swarm, ultrapilot, ultrawork, pipeline, and ecomode. The current primitives are ralph (iterative), team (parallel), and ccg (split-model).
ralph works through a PRD story by story, pausing at each for verification before advancing. It is the primary agent for sustained feature development. Between iterations, ralph polls the peer task queue so high-priority work can interrupt without manual intervention.
team spins up multiple tmux-based Claude Code sessions with explicitly defined file ownership per workstream. File ownership prevents conflicts. A session that does not own a file cannot write to it. Use team when you have three or more independent tasks (different files, different concerns) each taking 30+ minutes.
ccg routes backend work to Codex and frontend work to Gemini. Both run on their own token budgets, not your Anthropic quota. Use ccg when a feature requires API changes and UI changes that can happen in parallel and you have a clear contract between them before starting.
OMC's code-review agent fires automatically after any phase marked complete. It checks for logic errors, security issues, empty event handlers, and unresolved TODOs in a consistent format. A phase with an unresolved finding should not advance. Invoke it manually any time with the command below.
How do you know it worked?
If something looks off
Skills & Commands Reference
OMC uses magic keywords typed inline. No slash-command prefix needed. Every keyword triggers a different orchestration strategy with different model routing. 19 specialized agents auto-activate; pick by task shape.
| Command / Skill | What it does | When to use it |
|---|---|---|
teambasic | Canonical staged pipeline: plan → PRD → execute → verify → fix across coordinated Claude agents | Any feature that needs design, implementation, and verification, this is the default |
omc team N:codex/gemini/claudebasic | Spawn N real tmux CLI workers (actual Codex or Gemini processes in split panes) | When you need Codex for security/architecture review or Gemini for large-context UI work alongside Claude |
ccgbasic | Tri-model advisor: runs ask-codex + ask-gemini then Claude synthesizes both perspectives | Mixed backend+UI decisions where you want multiple model opinions before acting |
ralphbasic | Persistent execution with automatic verify/fix loops, won't stop until the task is confirmed complete | Tasks that must finish fully, no silent partials. Use for any critical feature or bug fix |
autopilotbasic | Single lead agent, fully autonomous end-to-end execution, minimal ceremony | Well-defined tasks where you don't need staged coordination, faster than team for simple features |
ulwadvanced | Maximum parallelism, burst parallel fixes/refactors without Team overhead | Many independent small fixes (lint errors, type errors, rename operations) that can run concurrently |
ralplanadvanced | Iterative planning consensus, multiple planning passes until a high-confidence plan emerges | Ambiguous or high-stakes tasks where you need the plan to be solid before any implementation starts |
deep-interviewbasic | Socratic requirements clarification, Claude asks structured questions to extract real requirements | Vague ideas or underspecified requests before starting any implementation |
deepsearchbasic | Routes to codebase-focused search, finds files, symbols, patterns across the repo | You know something exists in the codebase but don't know where it is |
ultrathinkadvanced | Deep extended reasoning mode for architectural and design problems | Complex architecture decisions, hard trade-offs, or when previous approaches have failed |
omc ask claude/codex/geminiadvanced | Run local provider CLIs and save a markdown artifact under .omc/artifacts/ask/ | Asynchronous second opinions or analysis you want saved for later reference |
omc waitbasic | Auto-resume daemon that waits for rate limit reset and continues your session | When you hit rate limits mid-session and want Claude to pick up automatically |
omc cost daily/weekly/monthlybasic | Usage and cost reports across all OMC sessions | Monitoring token spend or understanding which workflows cost the most |
/omc-setupbasic | Install or refresh all OMC configuration, agents, and skills | First install, after updates, or when OMC behavior seems off |
/omc-doctorbasic | Diagnose OMC configuration issues and report what's wrong | Something isn't working as expected, run this before filing a bug report |