🤖 agent-swarm

One goal. Six specialist agents. Your AI's own credit. A single prompt that turns any AI assistant into a parallel swarm — a planner, designer, frontend, backend, tester, and reviewer all working at once. No external CLIs. No API keys. Nothing to install on the AI's side.

What is agent-swarm?

agent-swarm is a prompt generator. You give it a goal, and it prints one self-contained prompt. Paste that prompt into any AI assistant — ZCode, ChatGPT, Claude Code, Codex, Gemini — and that assistant spawns its own 6 specialist subagents natively, in parallel, on its own account.

No CLI is launched. No API is called. The swarm runs inside the AI you already use, on the credit you already have.

flowchart LR
    G["🎯 Your goal"] --> O["🧠 agent-swarm<br/>(prints one prompt)"]
    O --> P["📋 Planner"]
    O --> D["🎨 Designer"]
    O --> F["🖥️ Frontend"]
    O --> B["⚙️ Backend"]
    O --> T["🧪 Tester"]
    O --> R["🔍 Reviewer"]
    P & D & F & B & T & R --> S["✅ Status board"]

Why this exists

Claude Code can spawn parallel subagents — but that pattern is locked to Claude Code. agent-swarm brings the same pattern to every AI assistant, with no CLI and no API integration. It's just a well-crafted instruction.

The core idea: the AI spawns its own subagents on its own credit. No external process can borrow another app's login — so agent-swarm doesn't try. It simply tells the AI to do it natively.

Install

npm install -g @hosnainrafi/agent-swarm

Or run once without installing:

npx @hosnainrafi/agent-swarm "build a dark-mode todo app"

Usage

agent-swarm "build a dark-mode todo app"     # default — print the swarm prompt
agent-swarm swarm "build a REST API"         # explicit form
agent-swarm zcode "build a todo app"         # write 6 ZCode subagents + launch prompt
agent-swarm teams                            # show the 6 specialist roles
agent-swarm help                             # help

Copy the printed prompt into any AI assistant. That's the whole workflow.

The prompt it generates

You are about to run a multi-agent swarm using YOUR OWN native subagents
(your built-in agent / subagent feature). Do NOT call any external API, CLI, or
third-party tool — every specialist runs inside you, on your own credit/session.

GOAL: build a dark-mode todo app

Spawn these 6 specialist agents and run them in parallel (simultaneously):

1. Planner   — concise implementation plan (scope, steps, acceptance criteria)
2. Designer  — design spec (tech stack, file structure, visual direction)
3. Frontend  — implement the frontend (write working files)
4. Backend   — implement the backend / data layer (write working files)
5. Tester    — test checklist, then run tests against the built code
6. Reviewer  — audit correctness, security, quality; score 0-10

Rules:
- Each specialist is YOUR OWN subagent on YOUR OWN credit. No external API.
- Run them concurrently. Each records its start time (HH:MM:SS) as its FIRST line.
- When all 6 finish, print a status board: agent -> start time -> one-line result.
- Then assemble and deliver the final result.

Do all of this now.

The 6 specialists

| Role | What it does | |---|---| | 📋 Planner | Scope, steps, acceptance criteria | | 🎨 Designer | Tech stack, structure, visual direction | | 🖥️ Frontend Dev | Implements the frontend | | ⚙️ Backend Dev | Implements the backend / data layer | | 🧪 Tester | Test checklist + runs tests | | 🔍 Reviewer | Correctness, security, quality (0–10) |

ZCode (Z.ai / GLM)

ZCode is a desktop app, not a CLI — so no Node package can launch it from the command line. What agent-swarm zcode does is write the 6 specialist subagents into ZCode's own directory (~/.zcode/agents/), then print a one-line launch prompt for you to paste into the ZCode app.

agent-swarm zcode "build a dark-mode todo app"

• Writes planner.md, designer.md, frontend.md, backend.md, tester.md, reviewer.md to ~/.zcode/agents/.
• Prints a launch prompt → paste it into ZCode, and ZCode runs the 6 subagents on your GLM credit.

Set ZCODE_AGENTS_DIR to override where the files are written.

FAQ

Does it use my AI's own credit? Yes. The prompt tells the AI to spawn subagents natively — same account, same credit, no new keys.

Does agent-swarm call any API or launch any CLI? No. It only prints text. The multi-agent work happens inside the AI you paste the prompt into.

Which assistants does it work with? Any assistant with a native subagent / agent feature: ZCode, ChatGPT, Claude Code, Codex CLI, Gemini, and more.

How do I know the agents ran in parallel? Each specialist records its start time (HH:MM:SS) as its first output line. If they cluster within seconds, they ran in parallel.

Support & contributing

⭐ Star the repo if it saves you time. Pull requests welcome — open an issue to report a bug or suggest an improvement.

License

MIT © Hosnain Rafi