AI team debates in your terminal. One question, multiple expert personas, structured discussion, actionable conclusion.
Instead of asking one AI for an answer, TeamTalk runs a full team discussion — Developer, Designer, PM, and Security Engineer each bring their perspective, challenge each other, and converge on a decision.
A single AI gives you one perspective. A team of AI personas gives you:
- Blind spots exposed — Security catches what Developer missed
- Real trade-offs — PM pushes back on Designer's idealism
- Better decisions — 3 rounds of debate, not 1 shot
Based on MIT's Society of Mind research — multi-agent debate improves reasoning by 15%+.
# One-liner (requires Go 1.22+)
go install github.com/Higangssh/teamtalk@latestOr build from source:
git clone https://github.com/Higangssh/teamtalk.git
cd teamtalk
go build -o teamtalk .# 1. Install
go install github.com/Higangssh/teamtalk@latest
# 2. Try it instantly (no API key needed)
teamtalk --demo
# 3. Use with real AI
export ANTHROPIC_API_KEY=sk-ant-... # or OPENAI_API_KEY=sk-...
teamtalk "Should we rewrite our monolith to microservices?"teamtalk "Do we need Kubernetes for 1000 users?"
teamtalk "Should we hire a junior or senior developer?"
teamtalk "Build vs buy for our auth system?"
teamtalk "Is GraphQL worth the complexity over REST?"Round 1 — Initial Opinions Each persona gives their take
Round 2 — Rebuttals They challenge each other's points
Round 3 — Final Positions Converge on a decision
Summary — Actionable conclusion with key agreements/disagreements
Default team:
| Persona | Role | Style |
|---|---|---|
| 💻 Developer | Technical feasibility | Blunt, hates complexity |
| 🎨 Designer | User experience | Fights for users |
| 📊 PM | Business impact & ROI | Pragmatic, data-driven |
| 🔒 Security | Risk & compliance | Paranoid, blocks risks |
| Provider | Model | Cost per debate |
|---|---|---|
| Anthropic | claude-3-haiku | ~$0.003 |
| Anthropic | claude-sonnet-4 | ~$0.03 |
| OpenAI | gpt-4o-mini | ~$0.003 |
| OpenAI | gpt-4o | ~$0.05 |
Override model:
TEAMTALK_MODEL=claude-sonnet-4-20250514 teamtalk "your question"Token usage is displayed after every debate:
📊 Token Usage
─────────────────────────────────
Provider: Anthropic
Model: claude-3-haiku-20240307
Input: 5,880 tokens
Output: 898 tokens
Total: 6,778 tokens
Cost: $0.0026
Single file. No frameworks. No dependencies.
main.go (514 lines)
├── Provider interface (Anthropic / OpenAI)
├── Debate engine (rounds, prompts, parallel calls)
├── Cost tracker (tokens, pricing per model)
├── Demo mode (built-in scenario, no API needed)
└── Terminal UI (typewriter effect)
- Custom personas via YAML (
--team team.yaml) - Ollama support (free local models)
- Streaming responses
- Export debate to Markdown
- MCP server mode
- TUI dashboard with Bubble Tea
MIT