oh-my-codex: Multi-Agent Orchestration Comes to Codex CLI

By Prahlad Menon 5 min read

If you’ve used oh-my-claudecode and wished the same multi-agent orchestration existed for OpenAI’s Codex CLI, Yeachan Heo built exactly that. oh-my-codex (OMX) is a pure add-on that layers 30 specialized agent personas and 39 skill workflows on top of Codex CLI — without forking it.

That “pure add-on” part matters. OMX uses native Codex CLI extension points, meaning it tracks upstream automatically. No merge conflicts, no version pinning, no maintenance burden when OpenAI ships updates:

npm install -g @openai/codex oh-my-codex
omx --madmax --high

What Are oh-my-codex Agent Personas and How Do They Work?

OMX ships 30 specialized personas: architect, executor, security reviewer, debugger, designer, and more. These aren’t prompt templates you swap manually — they’re orchestrated agents that hand off work to each other in structured pipelines, with up to 6 running in parallel.

Think of it like a development team where each member has a specific role. The architect agent acts as the quality gate, reviewing output from other agents and either signing off or sending work back for revision. This auto-verification loop catches the common failure mode where an AI agent confidently produces broken code and moves on — something any developer who’s used AI coding assistants has experienced.

What Are the Best oh-my-codex Workflows for Development?

OMX’s 39 skill workflows are where theory meets practice. Here are the five that matter most:

$deep-interview starts any project with a clarification phase. Instead of guessing at ambiguous requirements, OMX asks targeted questions to nail down scope, constraints, and edge cases before writing a single line of code. This front-loads the thinking that developers usually do mid-implementation — catching misunderstandings early when they’re cheap to fix.

$ralplan generates an implementation plan with explicit tradeoff analysis. It surfaces architectural decisions — database choice, API design, error handling strategy — and presents options with pros and cons before committing to a direction. You approve the plan before any code gets written.

$ralph is the execution engine: “don’t stop until done.” Once a plan is approved, $ralph drives implementation through to completion — file creation, test writing, and iterating on failures without manual intervention.

$team coordinates parallel agents on different parts of a codebase. Need frontend, backend, and tests written simultaneously? $team splits the work, assigns it to appropriate agent personas, and merges the results. This is where the 6-agent parallelism shines.

$autopilot chains everything end-to-end: interview → plan → execute → verify. Point it at a feature request and walk away. The architect agent reviews the final output before declaring it done.

Does oh-my-codex Remember Context Between Sessions?

Yes. OMX persists project context in a .omx/ directory, meaning agents pick up where they left off across sessions. Architecture decisions, coding conventions, known gotchas — all survive between runs. This is the difference between an agent that starts fresh every time (re-reading your entire codebase) and one that accumulates project knowledge over time.

This also helps with team workflows — when multiple developers use OMX on the same project, the .omx/ directory can be committed to the repo so everyone’s AI agents share the same project context.

Do I Need to Fork Codex CLI to Use oh-my-codex?

No, and this is a deliberate architectural choice. Forks create maintenance debt — every upstream change requires a merge, and eventually the fork either falls behind or becomes its own divergent project. OMX sidesteps this by treating Codex CLI as infrastructure and building orchestration on top via native extension points.

This means OMX benefits immediately from any Codex CLI improvements — new model support, performance optimizations, bug fixes — without work from the OMX maintainers. The approach mirrors what oh-my-claudecode did for Claude Code, and OMX claims ~92% feature parity with that project. The remaining 8% reflects differences between the underlying CLIs, not missing OMX functionality.

What’s the Difference Between oh-my-codex and oh-my-claudecode?

OMX is the Codex CLI equivalent of oh-my-claudecode (which does the same thing for Claude Code). Both provide multi-agent orchestration, specialized personas, and structured workflows. OMX claims ~92% feature parity. If you’re already in the OpenAI ecosystem using Codex CLI, OMX is the natural choice; if you use Claude Code, oh-my-claudecode is the equivalent tool.

Who Should Use oh-my-codex?

For individual developers: OMX turns Codex CLI from a chat-with-code tool into a structured development workflow. The $deep-interview → $ralplan → $ralph pipeline alone can save hours on feature implementation by front-loading architectural decisions.

For development teams: The $team workflow with parallel agents means you can prototype features faster. The architect agent’s auto-verification provides a quality baseline before human review.

For engineering managers: Project memory in .omx/ means onboarding new team members (or new AI sessions) gets faster over time as project context accumulates. Coding standards and architectural decisions are documented by the workflow itself.

OMX is recommended for macOS and Linux; Windows support exists but is second-class. The project is free and MIT licensed. You’ll need an OpenAI API key for Codex CLI, which has its own usage-based pricing.

npm install -g @openai/codex oh-my-codex
omx --madmax --high

Start with $deep-interview on your next feature to see the difference structured AI workflows make.