Everyone is talking about AI agents. Every week a new project goes viral on GitHub and promises to change how you work. It’s hard to tell what’s useful and what’s just noise.
Four names keep coming up: Hermes Agent, Everything Claude Code (ECC), oh-my-claudecode (OMC) and the Arena skill. People compare them as if you have to pick one. You don’t. They do completely different jobs. Here’s what each one is, in plain words.
01First, five words you’ll keep seeing
- AI agent
- An AI that doesn’t just answer. It takes actions: runs commands, edits files, sends messages, then checks the result and keeps going.
- Claude Code
- An AI agent for programming. It works inside your project, on your computer.
- Skill
- A saved set of instructions that teaches an agent one job, like a recipe card it can pull out when needed.
- Sub-agent
- A helper the main agent starts for one task. Like a manager handing work to a teammate.
- Tokens
- How AI usage is measured and billed. More work from the AI means more tokens, which means a bigger bill.
02Meet the four tools
Hermes Agent
A personal AI assistant that runs on its own.
You install it on your computer or a small cloud server, and it keeps working when your laptop is closed. You message it from Telegram, Slack, WhatsApp or Discord. It remembers you between conversations and can run tasks on a schedule. It works with many AI models, so you’re not tied to one company.
Think of it as: hiring an assistant who works from their own office and texts you updates.
ECC (Everything Claude Code)
An upgrade pack for an AI coding agent you already use.
It isn’t an agent by itself. It adds ready-made helpers, skills, rules and automatic checks to tools like Claude Code, Codex, Cursor and OpenCode. For example: a helper that plans before coding, one that reviews code, one that checks security, and one that makes the AI write tests first.
Think of it as: giving the assistant you already have a full training course and a professional toolkit.
oh-my-claudecode (OMC)
An add-on that runs Claude Code like a team.
Instead of one AI doing everything step by step, OMC splits a big task between several specialised agents, has them plan together, and keeps going until the result is checked, not just “looks done”. It also sends simple steps to cheaper AI models and hard ones to stronger models. You talk to it in plain words or with commands like /team and /execute.
Think of it as: a project manager who splits the work across a team and doesn’t sign off until it’s really finished.
Arena skill
One skill that makes many AIs compete for the best answer.
When Claude gives you a weak answer, you type /arena. It sends the same task to up to 100 helpers, and each one tries a different approach. They criticise each other’s answers in pairs, a judge scores each match, and the losers drop out round by round until one answer is left.
Think of it as: asking 100 people the same question, letting them debate, and keeping the answer that wins.
03Side by side
| Hermes Agent | ECC | OMC | Arena | |
|---|---|---|---|---|
| What it is | A complete AI assistant | Add-ons for a coding agent | Add-on that runs Claude Code as a team | One add-on skill |
| Who it’s for | Anyone who wants an assistant that works on its own | Developers and QA engineers | Developers with big, multi-step tasks | Claude Code users with one hard problem |
| What you need first | Nothing. Just an AI model account | Claude Code, Codex, Cursor or OpenCode | Claude Code | Claude Code |
| Use it | Every day, in the background | Every day, while you code | For big tasks that need a plan and several steps | Rarely, for important problems |
| Costs | Free software. You pay for AI usage, and a server if you host it | Free. Install only what you use, or it slows things down | Free. Several agents at once use more tokens, but it sends simple steps to cheaper models | Free, but uses a lot of tokens: about 600 AI calls for one full run |
| Be careful | It acts on its own and remembers what it learns about you | Its automatic checks run commands on your computer | Several agents multiply the bill. Don’t run its loops and another tool’s loops on the same task | The token bill. Try the small version first |
All four are free and open source (MIT license). On a phone, swipe the table sideways.
04Which one do I need, and why?
Start from your problem, not from the hype. Find your situation in the left column. The right column tells you why that tool fits and not the others.
| If you want… | Use | Why this one |
|---|---|---|
| An assistant you can text from your phone, that remembers you and does tasks while you sleep | Hermes Agent | It’s the only one that runs on its own, on a server, and talks through chat apps. The other three only work while you have Claude Code open. |
| A daily report or weekly check sent to Slack automatically | Hermes Agent | It has a built-in scheduler and sends results to chat apps. No one has to start it. |
| Your coding AI to plan first, write tests and review its own work | ECC | It adds the ready-made specialists for this: a planner, a test-first guide and code and security reviewers. OMC organises who works; ECC knows how the work should be done. |
| The same coding rules and security checks for your whole team | ECC | Its rules and automatic checks apply in every session, so standards don’t depend on someone remembering to ask. |
| Several AI agents to split up a big task, plan it together and finish it | OMC (/team) | It’s built to run Claude Code as a team. ECC gives you specialists, but OMC is the one that coordinates them in parallel. |
| Claude to keep going until the task is really done and checked | OMC (/execute) | It loops through checking and fixing until the result is verified, instead of stopping at “looks done”. |
| A better answer to one hard question Claude keeps getting wrong | Arena (/arena --quick) | Many different approaches compete and criticise each other, so weak ideas drop out. Too expensive for everyday use; worth it for one question that matters. --quick uses 16 helpers instead of 100. |
| To ask AI a question now and then | None of them | A normal chat with ChatGPT or Claude is enough. Installing a framework for a question only adds cost. |
05Install only what you need
Every add-on you install loads its instructions into every Claude Code session, and those instructions cost tokens even when you don’t use them. So install only the tools the table above pointed you to.
The commands below go in your normal terminal, not inside Claude Code, so you can paste a whole block at once. (Inside Claude Code, /plugin commands must be typed one line at a time.)
ECC
oh-my-claudecode (OMC)
Then open Claude Code and run /omc-setup once.
Arena
Hermes Agent
Hermes isn’t a Claude Code add-on. It’s its own program, with its own installer. Read the script before you run it: piping a script from the internet into your shell runs it with your permissions.
Can I install everything at once?
Yes. This block installs all three Claude Code add-ons in one go:
But “can” isn’t “should”. All three together means hundreds of extra skills and agents loaded into every session, and two tools (ECC and OMC) that can both run long loops. Install everything only if you really use everything. Before keeping a plugin, check what it costs you:
It lists everything the plugin adds and its projected token cost. If you stop using a tool, claude plugin disable switches it off without uninstalling it.
06Can I use them together?
Yes. ECC, OMC and Arena all plug into Claude Code, and they fit together: ECC decides how the work is done well, OMC decides who does it and in what order, and Arena settles the rare question that really matters. Hermes Agent is a separate program that runs next to them, for tasks that should happen without you.
One rule when you combine them: only one tool runs a long loop at a time. If OMC is pushing a task to the finish, don’t start another tool’s loop on the same task. My full ECC + OMC setup is in How I test with Claude Code.
I use ECC and OMC with Claude Code for my QA work, and Arena only for the rare hard question. An AI with every add-on loaded is like a test suite where every test is marked “critical”: when everything is important, nothing is.
07Example: a new project for example.com
Here’s how the two work together on a real kind of job. A company, example.com, asks you to set up automated tests for its website. You’ll use ECC for the everyday work and Arena for the one hard decision.
example.com is a placeholder name. On a real project, use the client’s site, and only test what they’ve allowed you to test.
Step 1. Install both, once
Type these inside Claude Code. You only do this the first time.
/plugin marketplace add affaan-m/ECC /plugin install ecc@ecc /plugin marketplace add Jakeschincariol/arena-skill /plugin install arena-skill@arena-skill
Step 2. Create the project
In your terminal, make a folder for the project and start Claude Code in it.
mkdir example-com-tests cd example-com-tests git init claude
Step 3. Let ECC pick what you need
ECC looks at your project and suggests which of its parts fit. It’s only a suggestion, so nothing changes until you agree.
/ecc:project-init
Step 4. Plan before any code
Ask for a plan. ECC writes it step by step and waits for your OK before it touches anything.
/ecc:plan Set up a Playwright + TypeScript test suite for example.com. Start with the home page, the navigation and the contact form.
Step 5. Use Arena for the hard decision
Every project has one question that really matters. Here it’s “what do we test first?” Instead of trusting the first answer, let 16 helpers compete. --quick keeps the token bill small.
/arena-skill:arena --quick Which three user flows on example.com should we automate first, and why? Rank them by business risk.
You get the winning answer, the criticism it survived and how many rounds it took. Feed that answer back into the plan.
Step 6. Build the tests
Ask Claude to write the tests with ECC’s testing skill, and to prove they’re stable.
Use the e2e-testing skill to write tests for the three flows we chose. Run every test twice and tell me if any result changes.
Step 7. Review and check before you share
Ask ECC to review the code, then to run its full check: build, tests, security and a look over every change.
/ecc:code-review /ecc:verification-loop
Fix what they find, commit, and send example.com a link with a short note on what the tests cover and what they caught.
| Step | Tool | Why this tool |
|---|---|---|
| Pick the setup | ECC | Installs only what this project needs |
| Plan | ECC | A written plan you approve before any code |
| Decide what to test first | Arena | One important question, so it’s worth the extra tokens |
| Write the tests | ECC | Its testing skill follows proven Playwright patterns |
| Review and check | ECC | Catches problems before the client sees them |
08How I filter the noise
I test software for a living, so I treat a viral AI project like a new build: it’s not trusted until it’s tested. Before I install anything, I ask five questions:
- What is it? A whole agent, an add-on for an agent, or one skill?
- What problem does it solve for me? If I can’t say it in one sentence, I don’t need it yet.
- What can it touch? My files, my computer, my accounts, my messages?
- What does it cost to use? “Free” software can still mean a big AI bill.
- How will I know it helped? I try it on one real task and compare the result with and without it.
GitHub stars show that a project is popular. They don’t show that it solves your problem. Start with the problem, then pick the tool.
Sources
- Hermes Agent by Nous Research, on GitHub
- ECC (Everything Claude Code) by Affaan Mustafa, on GitHub
- oh-my-claudecode by Yeachan Heo, on GitHub
- Arena skill by Jake Schincariol, on GitHub