oh-my-claudecode explained: turn Claude Code into a small dev team.

oh-my-claudecode explained

Claude Code is already a capable coding partner. Out of the box, though, a single conversation does everything: it plans, writes the code, tests it and reviews it. That works for small jobs. On bigger ones, the same assistant that wrote the code grades its own homework, and important steps get skipped.

oh-my-claudecode (OMC for short) tackles that problem. It's a plugin that organises Claude Code into a team of specialists with clear jobs and checkpoints. Think of it as hiring a small dev shop inside your terminal.

01What oh-my-claudecode actually is

OMC is an open-source project created by Yeachan Heo and released under the MIT license, so it's free to use, including commercially. It installs as a Claude Code plugin and adds three things:

• Specialist agents. Separate roles such as architect, executor, code reviewer, security reviewer and test engineer, each with its own instructions.
• Workflows (called skills). Commands like /autopilot, /team and /deep-interview that chain those agents together for you.
• Guard rails. Verification loops, a status bar that shows what's running, and project memory that carries lessons between sessions.

It's popular and moves fast. The project site showed roughly 40,000 GitHub stars in October 2026, and the v5.7 releases landed within days of each other that month. One naming quirk: the npm package is called oh-my-claude-sisyphus, not oh-my-claudecode.

02How it works: specialists, not one generalist

The core idea is simple: give each kind of task to the role best suited for it, and to the right-sized model. Claude comes in tiers. Haiku is fast and cheap, Sonnet is the everyday workhorse, and Opus is the slower, pricier heavy thinker. OMC pins each of its agents to a tier.

The pattern is easy to read. Opus does the thinking that's expensive to get wrong, such as planning, architecture and reviews. Sonnet does most of the hands-on building. Haiku handles quick searches and docs. The planner never writes the code it plans, and the reviewer never reviews its own work. That separation is the whole point.

03The modes worth knowing

OMC has a lot of commands, and you only need a handful. The two questions that matter are how clear the task is and how big it is.

If you remember one thing, make it this: start fuzzy ideas with /deep-interview. It asks Socratic questions and tracks how clear the requirements are before a line of code exists. For anyone who thinks better out loud than in specs, that's the most useful feature in the package.

04Team mode: a pipeline that checks its own work

/team is the recommended way to run several agents at once. Rather than letting them loose, it runs a fixed sequence of stages:

The second stage is the one that matters most. Writing down what "done" means before building gives the verify stage something objective to check against. And because the fix loop has a cap, a stuck task fails loudly instead of burning through your usage limits.

05The numbers: what's claimed vs. what's proven

OMC's website and README make bold promises: features shipped "3–5× faster" and "30–50%" token savings from smart model routing. Those numbers get repeated a lot, so we checked what they're based on.

That doesn't make the claims false. Routing simple work to cheaper models really does save money compared with running everything on Opus. But your savings depend on your tasks, and running more agents can also mean using more tokens overall. Treat the numbers as a direction, not a promise, and watch your own usage for a week.

06Why this matters if you're a founder in the making

Most early founders don't have a tech lead to push back on a bad plan, a senior engineer to review their code, or a QA person to catch the bug before customers do. OMC doesn't replace those people, but it builds their habits into how your AI assistant works:

• Clarity before code. /deep-interview forces the "what are we actually building?" conversation that saves weeks of rework.
• A second opinion by default. Separate reviewer and security-reviewer agents look at the work with fresh eyes, which the agent that wrote it can't do.
• A clear finish line. Verify loops push back on the classic AI failure: code that looks finished and isn't.
• Lessons that stick. /skillify turns a hard-won fix into a reusable skill file, so the same mistake doesn't cost you twice.

The founder skill this builds is delegation. You describe outcomes, set the definition of done and check the evidence. That's also the job of running a real team later.

07The catches

• It moves fast and breaks things. Version 5.0 (August 2026) removed 14 skills and 7 commands outright, without aliases. Read the release notes before updating, and expect older tutorials to show outdated commands.
• It's a lot. Around 48 skills is plenty to learn. Ignore most of them. Four or five commands cover nearly everything.
• It hooks into every prompt. OMC listens for keywords such as "autopilot" or "ralph" in your messages and adds its own instructions. If that gets in the way, the DISABLE_OMC and OMC_SKIP_HOOKS environment variables switch it off.
• It writes local state. Plans, logs and session data go into a .omc/ folder in your project. Keep it out of git, except .omc/skills/ if you want to share skills with a team.
• It uses your usage limits faster. More agents means more tokens per task. If you're on a tight plan, start with single-agent modes.

08How to try it on a side project

You need Claude Code installed and signed in. Then, inside a Claude Code session, enter these two commands one at a time. Pasting them together fails:

/plugin marketplace add https://github.com/Yeachan-Heo/oh-my-claudecode
/plugin install oh-my-claudecode

Run the setup, then start with a clarifying interview on something real:

/omc-setup
/deep-interview "a waitlist page for my app with email signup"

Once the plan is clear, hand it off:

/autopilot "build the waitlist page we just specced"

Want to try parallel Team mode later? Turn on Claude Code's experimental agent teams by adding this to ~/.claude/settings.json:

{
  "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" }
}

Then run something like /team 3:executor "fix all TypeScript errors". If something breaks after an update, /omc-doctor diagnoses the install.

09Verdict

oh-my-claudecode is one of the most complete attempts to make Claude Code behave like a team instead of a solo freelancer. The plan, build, verify and fix structure is the real value. It's free and MIT-licensed, and its building blocks can be checked in the code.

The speed and cost numbers are marketing until someone publishes results, and the pace of change means you'll occasionally relearn a command. If you already ship with Claude Code, install it on a side project this week and measure it yourself. If you're brand new, learn Claude Code first. The extra power only pays off once you know what to ask for.

SourcesSources and further reading

• oh-my-claudecode on GitHub: README, agent definitions, AGENTS.md and the benchmark harness (checked against the v5.6.1 package).
• oh-my-claudecode release notes: v5.0.0 breaking changes and the October 2026 v5.7 releases.
• Official oh-my-claudecode website: star count, features and the speed claim.
• oh-my-claude-sisyphus on npm: the package behind the omc command.
• Claude Code documentation: installation, models and settings.

Facts were checked on October 20, 2026. Tools like this change quickly; if something here is out of date, tell us on X.

Your Average Tech Founder

Host of The relatable entrepreneurship podcast, recorded in Da Nang, Vietnam. Writes about the tools, repos and habits that help founders in the making build better. No guru energy.