Autonomous AI Teams That Ship Production-Ready Code – Or Spark Your Next Big Idea
Developers and founders often face a "blank page" problem or get bogged down in boilerplate code. Existing AI tools often lack repository context, leading to manual copy-pasting, context loss, and high subscription costs for "black box" agentic services.
AI Symphony is an open-source, agentic framework that orchestrates specialized AI teams (Product Manager, Developer, Reviewer) to autonomously plan, code, and review software features. It operates directly on your repository, creating real Pull Requests without manual intervention. It also features a "Business Mode" for rapid startup idea validation.
AI Symphony is built on a modular architecture designed for reliability and transparency.
- Orchestration Engine: Powered by
CrewAI, managing agent roles, task delegation, and context sharing. - LLM Backend: Uses
LiteLLMvia OpenRouter to support 100+ models (Claude, GPT, Gemini, Llama). - Workspace Manager: A custom context manager for ephemeral Git cloning and secure file operations.
- Custom Tools: Specialized tools for
FileRead,CodeWrite, andGitHubPRoperations.
See architecture.md for a deeper dive.
| Your Current Pain | AI Symphony's Solution |
|---|---|
| Copy-pasting AI code → context loss & bugs | Agents clone your repo, write real files, commit & open PRs |
| PRs from tools like Devin are $500+/month | Open-source, free with local LLMs – mix Claude, GPT-4, or Qwen |
| Business ideas die in notebooks | New: Business Mode – AI crew spars ideas into actionable plans |
| Rigid agent setups (one role per tool) | YAML-configurable crews: Swap roles, add agents, tweak prompts |
uv run python src/main.py "Create a Streamlit leaderboard for top volleyball players by points, with Plotly charts and dark mode"- Yields: Real GitHub Pull Request with files, tests, and docs.
uv run python src/main.py --mode business "Idea: Subscription-based AI coach for teams"- Yields: Markdown report with pitch deck outline, risk matrix, and 5-year projections.
📂 See samples/business_mode_example.md for a complete example output.
- Python 3.10+
uv(recommended) orpip- GitHub Personal Access Token (PAT)
- OpenRouter API Key
Caution
Security: GitHub PAT Scoping
When creating your GitHub Personal Access Token, use minimal permissions:
- ✅
repo(for private repos) orpublic_repo(for public repos only) - ❌ Do NOT grant
admin,delete_repo, orworkflowpermissions - ❌ Never commit your
.envfile or share your token
Treat your PAT like a password. Rotate it regularly.
Option 1: Clone & Run (Recommended)
git clone https://github.com/traali/AI-Symphony.git
cd AI-Symphony
cp .env.example .env # Add your keys
uv sync
uv run python src/main.py "Your idea here"Option 2: Install via pip
pip install git+https://github.com/traali/AI-Symphony.gitAI Symphony supports local LLMs through LiteLLM, enabling cost-free and fully private operation.
# 1. Install Ollama: https://ollama.ai
# 2. Pull a model
ollama pull qwen2.5:14b
# 3. Update your .env
OPENROUTER_API_KEY=ollama # Use 'ollama' as placeholder
LITELLM_MODEL=ollama/qwen2.5:14b# 1. Download LM Studio: https://lmstudio.ai
# 2. Load any GGUF model and start the local server
# 3. Update your .env
OPENROUTER_API_KEY=lm-studio # Placeholder
LITELLM_MODEL=openai/local-model
LITELLM_API_BASE=http://localhost:1234/v1Tip
For best results with code generation, use models with 14B+ parameters like qwen2.5:14b, codellama:34b, or deepseek-coder:33b.
The repository includes several test scripts to verify your setup:
# Verify LLM connectivity
uv run python test_llm.py
# Test workspace tools directly
uv run python test_tools_directly.py
# Run a minimal end-to-end test (creates a PR)
uv run python minimal_test.pyCI status is automatically verified on every push. See the badge above for current status.
We welcome contributions! Please see CONTRIBUTING.md for details on how to join the orchestra.