AI Talk β€” Multi-Agent Discussion

AI experts debate your research questions, challenge assumptions, and reach structured conclusions

Open source Β· MIT
❓ How does this work?

AI Talk creates a panel of AI experts that debate each other on any topic. Unlike asking one AI, multiple AIs with different roles challenge each other β€” producing better-tested conclusions.

πŸ“
1. Set Topic
Paste research question or meeting notes
πŸ€–
2. Pick Agents
Load a team or add custom roles
πŸ’¬
3. They Debate
Agents argue, challenge, cite evidence
🎯
4. Conclusion
Structured recommendation with dissenting views
Why not just use ChatGPT? β€” A single AI tends to agree with you (sycophancy). Here, different AIs with opposing roles argue with each other, find flaws you'd miss, and produce conclusions tested by genuine intellectual friction.

Try it in one click Pick a topic β†’ 3-4 AIs debate it live β†’ get a verdict

No setup, no API key. Each demo spins up a panel of AI experts (GPT-4o, Claude, Gemini) that argue, cite live web sources, and reach a structured conclusion β€” streamed token by token.

Cheaper models (~$0.20 per demo instead of ~$0.80)

Discussion Mode

Choose how agents should approach the topic. This shapes their behavior and output style.

Conversation Topic Required

The more context you provide, the better the discussion. Paste research questions, meeting notes, or full articles.

Templates:
πŸ’‘ Tip: More context = better discussion. Paste meeting notes, paper drafts, or data summaries.

AI Agents

Add 2+ agents with different roles, or load a preset team.

Load Team:
Add One:

βš–οΈ Judge AI & Budget

The Judge evaluates after each round whether the agents have reached a solid conclusion.

πŸ’Ύ Saved Conversations

Your conversations are automatically saved. Resume any conversation anytime.

Total Cost
$0.0000
of $1.00 budget
Turns / Round
0
Round 0
Elapsed Time
0:00
running
Budget Utilization $0.0000 / $1.00
Thinking…
No discussion running
Head to the Setup tab to pick a one-click demo or configure your own panel of AI experts.
How it works Β· Tech & Architecture Multi-provider orchestration Β· SSE streaming Β· LLM-as-judge loop Β· token/cost accounting
Multi-provider orchestration
  • One async loop drives OpenAI GPT-4o, Anthropic Claude & Google Gemini behind a unified interface
  • Provider-specific SDK quirks (message roles, tool schemas) normalized per adapter
  • Mix models per seat β€” assign any role to any provider
SSE streaming
  • Token-by-token output over Server-Sent Events β€” no polling, no WebSocket overhead
  • Typed event protocol: round_start, turn_start, search, message, judge_verdict, conclusion
  • Mid-stream checkpoint events make every run resumable
LLM-as-judge loop
  • A judge model scores each round and returns CONTINUE or CONCLUDE with a confidence value
  • Tunable strictness + β€œjudge every N rounds” + min-rounds floor gate early exits
  • Schema-validated JSON with a safe CONTINUE fallback on malformed output
Web-search tool calls
  • Agents invoke a web_search tool to ground claims in live evidence
  • Bounded to 3 calls per turn; results truncated to keep context lean
  • DuckDuckGo by default with an optional SerpAPI fallback
Token & cost accounting
  • Per-provider input/output tokens metered with tiktoken
  • Live USD cost from a per-model price table, enforced against a run budget
  • Auto-stops on budget_exceeded or max_turns_reached
Lean, dependency-light stack
  • Single-file FastAPI backend, single-file vanilla-JS SPA β€” no build step
  • State + history persisted in localStorage; transcripts export to Markdown
  • Accessible by design: skip link, live regions, reduced-motion support
Stack Python 3.12 FastAPI Uvicorn Β· SSE Vanilla JS OpenAI Β· Anthropic Β· Gemini SDKs tiktoken Docker Read the source