Claude Code project management layer (not an IDE replacement)
OMC first launched on 9 / 1 / 2026 (day / month / year), accumulating 35,069 stars and 3,213 forks within five months. Author Yeachan Heo describes it as Teams-first Multi-agent orchestration for Claude Code. It does not generate code directly; it builds a multi-agent framework above Claude Code that divides tasks, assigns work, runs tests, and verifies results automatically.
OMC is distributed as both an npm package (oh-my-claude-sisyphus) and a Claude Code Plugin. Installation adds 19 specialist agents, including architect, executor, qa-tester, security-reviewer, and debugger, plus 27 slash commands and the Team pipeline. Enter /autopilot "build a REST API for managing tasks" to have OMC handle planning, execution, verification, and correction automatically.
v4.4.0 removed the codex and gemini MCP servers in favor of real CLI worker pools opened directly through tmux (omc team 2:codex "..." opens two codex panes). codex and gemini start as external tools on demand and close after execution. This is OMC's concrete implementation of "team-first."
team-plan
team-prd
team-exec
team-verify
team-fix
Don't learn Claude Code. Just use OMC.
Plugin installation, setup, and autopilot startup
OMC provides two installation paths. The Claude Code Plugin marketplace is recommended for most users. Enter the following commands sequentially in a session; the README warns that pasting both lines at once fails, so send them separately.
# First command: add the marketplace source
/plugin marketplace add https://github.com/Yeachan-Heo/oh-my-claudecode
# Second command: install the plugin (wait for the previous command to finish before sending this)
/plugin install oh-my-claudecodenpm terminal installation
The npm package is named oh-my-claude-sisyphus: the repository is oh-my-claudecode, but npm retains the old name. Installation adds the omc terminal command for running subcommands such as team, ask, and wait directly from the shell.
npm i -g oh-my-claude-sisyphus@latestsetup and native Claude Code Teams configuration
After installation, enter /setup or /omc-setup in a Claude Code session, or run omc setup from the terminal. Team mode also requires enabling Claude Code's experimental Teams setting. Add the following configuration to ~/.claude/settings.json.
{
"env": {
"CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
}
}OMC's 20 orchestration modes and common commands
OMC separates orchestration into two layers: modes determine how tasks are assigned to multiple agents, and slash commands perform specific operations. The following cards cover the core 20 items listed in the README: the first 10 are modes and the last 10 are common commands. Use /team, /autopilot, and /ralph as the main entry points for everyday work.
Mode · 01
/team
canonical orchestration
Officially recommended mode after v4.1.7. Pipeline: plan → prd → exec → verify → fix loop. Requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS.
Mode · 02
omc team
tmux CLI workers
Opens tmux panes from the terminal for concurrent real claude, codex, and gemini CLI processes. Workers close after use and do not continue occupying resources.
Mode · 03
/ccg
tri-model advisor
Calls codex and gemini together in one command. Claude synthesizes both responses into the final conclusion.
Mode · 04
/autopilot
end-to-end autonomy
A single lead agent executes autonomously. Given 'Build X,' OMC plans, writes, executes, and verifies. The entry mode with the fewest operating steps.
Mode · 05
/ralph
persistent verify loop
Persistent verification mode. The verify/fix loop continues until all evidence passes without stopping early. Includes ultrawork parallelism.
Mode · 06
/ultrawork
maximum parallelism
High-concurrency correction mode outside team scenarios. Edits multiple files at once, skips PRD, and does not run verify automatically. Shorthand ulw.
Mode · 07
UltraQA
repeated quality gates
Continues until tests / build / lint / typecheck all pass. Handles long-tail errors in which fixing A breaks B.
Mode · 08
/ralplan
iterative plan consensus
ralph for planning. Iterates on the same plan until multiple agents reach consensus, then enters exec.
Mode · 09
Pipeline
sequential stages
Strictly ordered multi-stage transformations for ETL, migration, format conversion, and other tasks whose stages must remain in order.
Mode · 10
Ultragoal
artifact-only goals
Creates a goals + checkpoints + evidence file without starting a loop. Suited to handoffs, audits, and situations where loops cannot run.
Skill · 11
/deep-interview
socratic clarification
Socratic questioning. Quantifies the clarity of vague ideas through weighted dimensions, identifies hidden assumptions, and passes the result to a downstream mode.
Skill · 12
/ask
provider advisor
Sends the current context to other provider CLIs (claude / codex / gemini) for advice and saves it automatically as Markdown in .omc/artifacts/ask/.
Skill · 13
/skill
skill library
Lists, adds, deletes, edits, and searches your accumulated .omc/skills/*.md files. Project-level skills override user-level ones and are injected automatically when a trigger matches.
Skill · 14
/skillify
pattern extractor
Extracts the current session's problem-solving process into a reusable skill file after strict quality gates, for automatic injection in later tasks.
Skill · 15
/oh-my-claudecode:autoresearch
bounded research loop
The standard research workflow after v4.4. Start a mission with /deep-interview --autoresearch, then run a bounded stateful loop.
Skill · 16
/oh-my-claudecode:hud
live statusline
Live observation line. The focused preset displays the current agent, token usage, stage, and subtask count in the Claude Code status line.
Skill · 17
/omc-doctor
health check
Error diagnostics. Checks plugin paths, cache, CLI tools, tmux, and environment variables, and removes old plugin cache.
Skill · 18
omc wait
rate-limit resume
Session recovery during Claude rate limits. omc wait --start starts a background daemon that automatically restarts the original session once the allowance returns.
Skill · 19
ultrathink
deep reasoning
Keyword-triggered mode. Adding 'ultrathink about ...' to the prompt switches OMC to its deepest thinking path and higher-weight model reasoning.
Skill · 20
cancelomc / stopomc
emergency stop
Emergency stop commands for all OMC modes. Enter stopomc to terminate all running modes immediately when a loop malfunctions, autopilot deviates, or token usage grows too high.
Scenarios and recommended modes
| Scenario | Recommended mode | Reason |
|---|---|---|
| Multiple files, a PRD, and staged collaboration | /team | plan→prd→exec→verify→fix pipeline with native Claude Teams collaboration. |
| A one-sentence feature request with fewer operating steps | /autopilot | One lead agent runs the complete workflow with the lowest entry barrier. |
| No stopping before a fix, with evidence-based completion | /ralph | verify/fix loop and built-in ultrawork until all evidence passes. |
| Concurrent real codex / gemini CLIs | omc team N:codex "..." | tmux worker pools from v4.4 with real concurrent CLIs across models. |
| Unclear requirements | /deep-interview | Socratic questioning quantifies the specification before choosing a downstream mode. |
10 common configuration errors and fixes from official documentation
OMC's major versions change frequently, and lengthy README and documentation sections make important settings easy to miss. These 10 points come from the README, docs, CHANGELOG, and plugin structure. They are common first-installation issues often discovered only in the first week.
Send the two installation commands separately
The README states 'pasting both lines at once will fail.' Wait for /plugin marketplace add to finish before sending /plugin install; do not paste both together.
Source · README · Quick Start
The npm package is not named oh-my-claudecode
The repository, plugin, and commands are called oh-my-claudecode, but npm retains the old package name oh-my-claude-sisyphus. npm i -g oh-my-claudecode installs the wrong package. Use npm i -g oh-my-claude-sisyphus@latest.
Source · README · Package naming note
/team fallback without EXPERIMENTAL_AGENT_TEAMS
Team mode depends on Claude Code's experimental native Teams API. Without CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in ~/.claude/settings.json, OMC warns and falls back to non-team execution. It may appear to work but has lost pipeline functionality.
Source · README · Team Mode section
codex / gemini MCP server removal in v4.4
From v4.4.0, the x and g provider MCP servers are removed. Older MCP installation tutorials are obsolete. Use omc team 2:codex "..." to open real codex CLI panes in tmux.
Source · README · v4.4.0 release note
swarm removal and the canonical /team mode
From v4.1.7, the legacy swarm keyword and skill are removed entirely. Replace /swarm ... in old articles with /team .... /omc-teams is only a compatibility alias routed internally to omc team.
Source · README · Team Mode section
No CLI command for autopilot
The terminal does not support omc autopilot; the README explicitly states that no such subcommand exists. autopilot, ralph, ultrawork, and deep-interview are session-only and work only inside Claude Code chat.
Source · README · CLI vs In-session table
Commit skills to git for worktree sharing
Project-level skills live in .omc/skills/. If added in a worktree without a commit, they disappear when the worktree is deleted. Use git add + commit for long-term preservation. User-level skills in ~/.omc/skills/ are shared across projects.
Source · README · Custom Skills section
autoresearch is a hard-deprecated shim
The old omc autoresearch is hard-deprecated. The new workflow creates a mission with /deep-interview --autoresearch, then runs a stateful loop with /oh-my-claudecode:autoresearch. Skip old instructions using omc autoresearch.
Source · README · Utilities · Autoresearch
Automatic rate-limit recovery with omc wait
Long-running ralph or team sessions can encounter Claude rate limits. Start the background daemon with omc wait --start; it resumes the paused session automatically when the allowance returns. Session detection requires tmux.
Source · README · Rate Limit Wait section
Set OMC_PLUGIN_ROOT in --plugin-dir mode
When starting with claude --plugin-dir <path>, HUD and setup cannot find the corresponding plugin. Set export OMC_PLUGIN_ROOT=<path> in the shell or add --plugin-dir-mode during setup, otherwise plugin installation and runtime paths diverge.
Source · docs/REFERENCE.md · Plugin directory flags
/autopilot REST API task stages
This demonstrates execution of the README example /autopilot "build a REST API for managing tasks", showing OMC stage changes, HUD observation, and ralph's automatic correction loop when verification fails.
$ You › /autopilot "build a REST API for managing tasks"
claude: planner ›
First break the task into milestones:
M1 · schema + models (Task, User, Tag)
M2 · CRUD endpoints + validation
M3 · authentication + rate limit
M4 · tests + OpenAPI documentation
claude: architect ›
Recommended stack: Fastify + Prisma + Postgres, reasons:
· Native TS support, a clean async model
· Prisma is more intuitive for relational tables such as task
$ You ›
ok use this stack. Continue.
claude: executor ›
Wrote prisma/schema.prisma and src/models/task.ts
· 14 files created · 412 lines
claude: verifier ›
Run tests / build / typecheck / lint:
✓ build · ✓ typecheck · ✓ lint
hl: ✗ tests · 3/27 failing (task.search edge cases)
claude: autopilot ›
3 tests failed, shall I switch to ralph mode and fix them until all pass?
$ You › /ralph "fix the failing task.search tests"
claude: debugger › Root cause: An empty query string was not rejected. In the Prisma where condition, it
hl: triggered case-insensitive LIKE, bypassing the index and causing a timeout.
ok: [ralph exited · evidence ledger updated]
ok: · build ✓ · typecheck ✓ · lint ✓ · tests 27/27 ✓
$ You › /ccg "review the auth flow"
claude: claude (synth) ›
Both flagged the same issues: refresh token rotation is missing,
the client has no graceful fallback when the session expires.
codex also noted: the rate limiter is in-memory, and breaks during scale-out.
gemini also noted: error messages create a user enumeration risk.
$ You › /skillify
ok: [saved .omc/skills/fastify-prisma-task-api.md]
When autopilot detects a verification gap, it proactively suggests switching to ralph mode until all tests pass.
Three mode collaboration mechanisms
autopilot uses an orchestration layer, calling planner → architect → executor → verifier in order. Each agent has an independent SKILL prompt; handoffs carry the full context so the verifier need not reread files. ralph supplements autopilot with persistent verification: autopilot proactively suggests switching when it detects a verification gap, without manual judgment about entering a correction loop.
/skillify extracts the successful workflow into .omc/skills/*.md for automatic injection into similar later tasks. Project-level problem-solving patterns can be saved and reused, with updates faster than a model training cycle.
OMC functional boundaries and known limitations
OMC as a team's standard tool
After adapting skills, settings, and HUD to your codebase, new tasks have a higher verification pass rate than with Claude Code alone. These five paths turn OMC from an individual tool into a shared team resource.
Advanced configuration paths
1. Commit project-level skills to git. Commit .omc/skills/*.md to the project repository so new engineers receive historical problem-solving records after cloning. Trigger matches inject them automatically, keeping past solutions available for later tasks.
2. Configure the HUD focused preset. After /oh-my-claudecode:hud setup, add "omcHud": { "preset": "focused" } to ~/.claude/settings.json. The status line displays the current agent, stage, and token usage for progress monitoring during long runs.
3. Use /ccg for cross-CLI advice. For codex architecture review and gemini UX perspectives, use /ccg "review this PR". OMC queries both and Claude synthesizes the results, more reliably than switching windows manually.
4. Pair with oh-my-codex. The same author's sibling project oh-my-codex brings the same orchestration experience to OpenAI Codex CLI for team members who prefer codex.
5. Follow updates in the official Discord. OMC v4.x has undergone several major revisions. The official Discord usually provides workarounds 1–2 patches ahead of the issue tracker.
Three reference documents
① Official documentation: Full directory of CLI Reference, Workflows, and Execution Modes. ② docs/REFERENCE.md: All flags and the decision matrix for plugin-dir mode. ③ CHANGELOG.md: Complete record of breaking changes, including MCP removal in v4.4 and swarm deprecation in v4.1.
Your Claude Just Have been Steroided.