oh-my-codex

Your codex is not alone.

Multi-agent orchestration layer for OpenAI Codex CLI. Build faster with staged team pipelines, persistent memory/state MCP servers, and extensible hooks.

🚀 v0.21.3 Latest Release ★ — Stars ↓ — Downloads/mo ⚡ 32 Prompts 🎯 25 Skills 🤝 Team Mode 🔌 6 MCP Servers
Get Started Read Docs Vibetip Docs View on GitHub
Recommended Workflow 
# 1. Launch interactive Codex session
$ omx --madmax --high

# 2. Use keywords inside Codex for workflows
$deep-interview "clarify the auth change"
$ralplan "approve the implementation plan"
$ultragoal "carry the plan to completion"
Launch omx --madmax --high

Start the interactive Codex session with recommended settings for trusted environments.

Clarify $deep-interview

Use inside Codex when intent or boundaries are still unclear. Socratic clarification before planning.

Execute $ultragoal or $team

Choose $ultragoal for a durable goal ledger that runs to completion, or $team for coordinated parallel execution.

oh-my-codex character illustration

v0.21.3

Released: September 3, 2026 — duplicate Team wakes are eliminated by retiring terminal projections, the detached leader tmux scrollback clamp rises from 500 to 5000 lines, and composer prompt-drift triage is documented. The 0.21 line also retired the legacy execution loops behind a sunset-stub resolver and removed the MCP state writer tools.

Team Wake De-duplication

Team no longer re-wakes on already-terminal projections; terminal projections are retired instead of lingering, so a finished worker stops generating wake churn.

5000-Line Scrollback Clamp

The detached leader scrollback clamp was 500 lines, so long sessions silently discarded multi-line assistant responses. It now holds 5000 lines.

Sunset-Stub Resolver

Retired keywords resolve to a direct replacement instead of failing quietly: $ralph → $ultragoal, $ultrawork → $team, $pipeline → $plan + $team, $autoresearch-goal → $autoresearch.

Read-Only MCP State

The MCP state writer tools state_write and state_clear are removed from the surface and return explicit deprecation errors; durable writes go through the CLI.

Canonical omx autopilot

A first-class CLI command restores the staged $deep-interview → $ralplan flow, and the standalone autopilot/pipeline loops are merged into the unified planning surface.

$team — Worktree-First Parallel Execution

The core OMX skill. Since v0.13.1, every team worker runs in an isolated git worktree by default — no flags needed, no merge conflicts, automatic integration.

Automatic Worktree Isolation

Every worker gets a dedicated git worktree at .omx/team/<name>/worktrees/worker-N. No --worktree flag needed — it's the default. Workers write to isolated detached branches; the leader workspace stays clean.

Incremental Merge Tracking

The leader continuously integrates worker commits via merge, cherry-pick, or cross-worker rebase strategies. Conflicts are detected early and reported in integration-report.md. Auto-commit ensures no work is lost.

Mixed-Provider Teams

Run Codex, Claude, and Gemini workers side-by-side with OMX_TEAM_WORKER_CLI_MAP. Each worker gets its own worktree regardless of provider. Full staged pipeline: plan → exec → verify → fix.

$team — Worktree Lifecycle 
# Use $team inside Codex (recommended)
$team 3:executor "refactor auth module"
Team started: refactor-auth-module
workers: 3 (worktrees: automatic, detached)

# Advanced: CLI operator surface
omx team 3:executor "refactor auth module"
omx team status refactor-auth-module

# Mixed providers with worktrees
OMX_TEAM_WORKER_CLI_MAP=codex,claude,gemini \
  omx team 3:executor "full-stack implementation"

Features

Everything you need for autonomous software development with OpenAI Codex CLI

Role Prompts

32 specialized prompts via /prompts:name for targeted expertise across implementation, review, domain, and coordination lanes.

Workflow Skills

25 workflow skills via $name syntax for autopilot, team orchestration, planning, recovery, and QA cycles.

Team Orchestration

Isolated parallel execution with automatic git worktrees. Workers get dedicated worktrees by default — zero merge conflicts, automatic commit integration, and incremental merge tracking.

MCP Servers

State, memory, code intelligence, trace, wiki, and Hermes servers (6 total) for persistent context and cross-session learning.

Staged Pipeline

plan → prd → exec → verify → fix lifecycle for structured, evidence-based development workflows.

Launch Profiles

--yolo, --high, --xhigh, --madmax modes for granular control over reasoning effort and execution intensity.

Installation

Get up and running in seconds with npm

Step 1: Install via npm
npm install -g @openai/codex npm install -g oh-my-codex
Step 2: Setup and verify
omx setup && omx doctor
Step 3: Launch and work
omx --madmax --high

Agent Tiers

32 specialized prompts organized by domain, review depth, and orchestration responsibility

Build/Analysis

Core Development

architect planner executor debugger verifier explore explore-harness analyst code-simplifier
Review Lane

Quality Assurance

style-reviewer quality-reviewer api-reviewer security-reviewer performance-reviewer code-reviewer critic quality-strategist
Specialists

Domain Experts

dependency-expert test-engineer build-fixer designer writer qa-tester git-master researcher vision
Product

Product Strategy

product-manager ux-researcher product-analyst information-architect
Team Mode

Team Orchestration

team-orchestrator team-executor

Skills

Repeatable workflows triggered by keywords or $syntax

Keyword Skill Name Description
autopilot autopilot Full autonomous execution from idea to working code
worker worker Single worker lane driven by the team runtime
team team N coordinated agents using Codex CLI native teams
plan plan Strategic planning with consensus and review modes
ralplan ralplan Iterative planning consensus that ends on a reconciled receipt
ultragoal ultragoal Durable goal ledger that tracks every deliverable to completion
autoresearch autoresearch Goal-directed research mission that ends on a structured verdict
deep-interview deep-interview Requirements interview that produces a spec before any mutation
cancel cancel Stop execution modes and clear active state
design design Canonical product/UI decision workflow recorded in DESIGN.md
analyze analyze Repository analysis pass before planning or implementation
hud hud Live session HUD with ownership-scoped panes
code-review code-review Comprehensive code review across all quality dimensions
ultraqa ultraqa Adversarial QA with end-to-end expectations
wiki wiki Repository wiki generation and lookup
doctor doctor Diagnose and repair a broken OMX runtime or plugin install
ralph — Retired in 0.21 — use ultragoal
ulw / ultrawork — Retired in 0.21 — use team
pipeline — Retired in 0.21 — use plan and team
autoresearch-goal — Retired in 0.21 — use autoresearch

Discord Integration

Seamlessly connect your Codex workflows with Discord for real-time notifications and team collaboration.

Webhook Integration

Easily send build status, test results, and agent updates to your Discord channels using the configure-notifications skill.

Skill
"configure discord webhook"

Bot Interaction

Deploy custom bots to listen for commands, trigger workflows, and provide interactive feedback directly within your Discord server.

Skill
"setup discord bot"

OpenClaw Gateway

Unified outbound delivery for all your notification channels with intelligent routing and status signaling.

Unified Delivery

Send notifications to Discord, Slack, Telegram, and custom webhooks through a single OpenClaw gateway. Intelligent routing ensures messages reach the right channel.

Status Signaling

Reliable delivery confirmation with retry logic. Track notification status in real-time and ensure critical alerts never get lost.

Telegram Integration

Connect your Codex workflows to Telegram for instant mobile notifications and bot interactions.

Bot Notifications

Configure your Telegram bot to receive build updates, deployment notifications, and agent status messages directly in your chat.

Skill
"configure telegram notifications"

Command Interface

Control your Codex agents via Telegram messages. Start workflows, check status, and receive summaries without leaving your mobile device.

Skill
"setup telegram commands"
GEO visibility

Measure whether LLMs can find OmX

oh-my-codex ships a geobench product spec for tracking hit rate, MRR, share of voice, and citations across LLM answer surfaces.

Product spec

YAML input for geobench profiling and benchmark runs.

Runbook

How to estimate, profile, run, and publish GEO metrics safely.