ai-team-builder — Your AI Software Company in the Terminal

ai-team-builder is a model-agnostic, multi-agent AI coding system. Give it a single sentence describing your idea, and a team of specialized AI agents builds the software for you:

| Agent | Role | | --- | --- | | CEO (Planner) | Analyzes your idea and writes a full product specification | | Designer (Architect) | Picks the tech stack and designs the file structure | | Developer | Writes complete, production-quality code | | Tester | Actually runs your code and captures real errors | | BugFixer | Diagnoses real test failures and fixes the code | | Reviewer | Audits quality, security, and spec compliance, scores 0–10 | | Writer | Produces a complete README manual |

This is inspired by ChatDev (the famous "virtual software company") but enhanced in the ways that matter:

1. Real execution loop — ChatDev only talked about code; ai-team-builder's Tester actually runs the generated code and feeds real error output to the BugFixer. Bugs are fixed against real failures, not guesses.
2. Model-agnostic — works with GPT / Codex (OpenAI), GLM (Zhipu), Gemini (Google), DeepSeek, Qwen, Claude, OpenRouter (hundreds of models), and local Ollama models. Any OpenAI-compatible API works — usually a one-line env change.
3. No API key required — run it 100% free with local Ollama models: no account, no quota, no internet. Just install Ollama, pull a model, and go.
4. Per-role model mixing — use a powerful model for planning/review and a cheap model for coding/testing: --planner-model gpt-5 --model gpt-5-nano.
5. Zero-install npx usage — npx ai-team-builder "your idea" builds software with nothing installed.
6. Full transparency — every agent conversation is logged to ai-team.session.log and ai-team.transcript.json, and a review.md + test-report.md ship with the project.
7. Sandboxed execution — dangerous commands are blocked; runs are timed out.

Quick Start

# Run once without installing
npx ai-team-builder build "a todo app with dark mode saved to localStorage"

# Or install globally
npm install -g ai-team-builder
ai-team-builder build "snake game with score tracking"

Use it completely free — no API key, no account

# 1. Install Ollama: https://ollama.com
# 2. Pull a coder model
ollama pull qwen3-coder:30b
# 3. That's it — ai-team-builder detects Ollama automatically and uses it
ai-team-builder build "a markdown-to-HTML converter CLI"

If no API key is set, ai-team-builder automatically falls back to your local Ollama. Bring your own key whenever you want cloud models.

Setup for your preferred cloud AI model (optional)

# OpenAI GPT / Codex (default when a key is present)
export OPENAI_API_KEY=sk-...

# Zhipu GLM
export GLM_API_KEY=...
export AI_TEAM_PROVIDER=glm

# Google Gemini (OpenAI-compatible mode)
export GEMINI_API_KEY=...
export AI_TEAM_PROVIDER=gemini

# DeepSeek
export DEEPSEEK_API_KEY=...
export AI_TEAM_PROVIDER=deepseek

# Alibaba Qwen
export QWEN_API_KEY=...
export AI_TEAM_PROVIDER=qwen

# OpenRouter (choose from 100+ models incl. Claude, GLM, Gemini, Llama)
export OPENROUTER_API_KEY=...
export AI_TEAM_PROVIDER=openrouter

# Local Ollama (explicit, e.g. a smaller model)
export AI_TEAM_PROVIDER=ollama
export AI_TEAM_MODEL=qwen3-coder

# Any custom OpenAI-compatible endpoint (Azure, LM Studio, vLLM, litellm...)
export OPENAI_BASE_URL=http://localhost:11434/v1
export OPENAI_API_KEY=ollama

See all presets and which keys are set: ai-team-builder models

CLI Reference

ai-team-builder build "<idea>"     Build software from an idea (default command)
ai-team-builder "<idea>"           Same as build
ai-team-builder models             Show provider presets and configured keys
ai-team-builder help               Show help

Options:
  --name <dir>             Output directory (default: auto-generated)
  --provider <preset>      openai | codex | glm | gemini | deepseek | qwen | openrouter | anthropic | ollama | builtin
  --model <model>          Model name, e.g. gpt-5.5, glm-4.7, gemini-2.5-pro, qwen3-coder
  --planner-model <m>      Different model for CEO/Reviewer roles
  --max-loops <n>          Max test-fix cycles (default 4)
  --yes                    Skip interactive confirmations
  --quiet                  Only show the final summary

What you get

After a run, your project directory contains:

my-project/
├── <generated source files>
├── README.md              # usage manual written by the Writer agent
├── design.md              # the Designer's architecture plan
├── review.md              # Reviewer's score + issues
├── test-report.md         # which test commands passed
├── ai-team.session.log    # full agent chat replay
└── ai-team.transcript.json # structured session transcript

How the pipeline works

1. CEO turns your idea into a spec → 2. Designer plans stack + files → 3. Developer writes all code → 4. Tester runs real commands; on failure the BugFixer diagnoses from the actual error output and patches the files → (repeat up to 4 cycles) → 5. Reviewer audits everything and grades it → 6. Writer documents the project.

Per-project config

Drop a .aiteam.json in your project folder:

{
  "provider": "glm",
  "apiKey": "...",
  "model": "glm-4.7",
  "roleModels": { "ceo": "glm-4.7", "developer": "glm-4-flash" }
}

FAQ

Can it run without any API key? Yes. Install Ollama (free, open source), pull any coder model, and ai-team-builder uses your local models automatically — no account, no quota, no internet connection required after pulling the model.

Which models work best? Frontier models (GPT-5.x, Claude Sonnet/Opus, GLM-4.7+, Gemini 2.5 Pro) give the best code quality. Smaller/cheaper models (including local Ollama models) work for planning and simple projects.

Why is this better than one-shot "write me code" prompts? Specialization + a real test-fix loop. Each agent has a focused role prompt, and generated code is executed and debugged against real errors — the two biggest weaknesses of single-prompt code generation.

Why does this tool use API keys at all? The AI models themselves run on other companies' servers (OpenAI, Google, Zhipu…), so each user normally brings their own key. ai-team-builder minimizes this friction: any OpenAI-compatible provider works, keys are picked up from env vars, and the Ollama path removes the need entirely.

Is it safe? Commands execute in the output directory only, with a 30–60s timeout, and a blocklist of destructive operations (rm -rf /, curl, wget, package installs, interactive shells, etc.).

License: MIT.