# LLM Debate Council

> A council of AI models from different vendors (OpenAI GPT, Anthropic Claude, Google Gemini,
> xAI Grok, DeepSeek, Qwen) that **debate the user's question**, search the web and fact-check
> themselves, draw on the user's own uploaded knowledge base, and converge on one well-reasoned
> decision — with a confidence score, the points of disagreement, and cited sources.

Site: https://llmd.cc

## What it is
Instead of one AI giving one answer, several models from different companies argue a question
from every angle, challenge each other, verify facts on the web, and reach a single decision the
user can defend. Different vendors have different strengths, so the council catches blind spots a
single model would miss. Two ways to run it: built-in agents with included monthly credits
(Free: 3 modes and 5 debates/month; Starter/Pro subscriptions unlock more modes, chat and bigger
credit pools) — or bring your own API keys, with tokens billed on your provider account, no markup.

## Who it's for
Founders, product managers, analysts, researchers and consultants who need a verified,
multi-perspective answer for an important decision — not just one model's opinion.

## How it works (in 4 steps)
1. You pick 2–10 AI agents (models from any supported vendors) and type your question or idea.
   One-click role presets (Finance, Legal, Risk, Strategy, Marketing, Engineering, Operations,
   Devil's Advocate) build an expert panel fast.
2. The agents answer in rounds, building on or challenging each other. If web access is on they
   search the internet first and ground their arguments in real sources. Relevant excerpts from
   your personal knowledge base are pulled in automatically.
3. Optionally a verifier fact-checks the claims on the web (VERIFIED / DISPUTED / FALSE with links),
   and an optional Red-team pass attacks the emerging conclusion from every angle before the final.
4. A final synthesis gives ONE decision with the key reasoning, a **confidence score**, a
   **"where they disagreed"** breakdown, and the **list of sources** used — exportable to PDF/DOCX.

## Eleven thinking modes
- Debate — sides argue to consensus, with agreement/confidence metrics and a verifier.
- Brainstorm — maximum ideas without criticism; stops when no genuinely new ideas appear.
- Idea development — take one idea and deepen it step by step.
- Red Team — a defender vs attackers probing weaknesses by angle (security, scale, cost, UX, legal).
- Role analysis — each agent plays a role from a framework: SWOT / Six Thinking Hats / Stakeholders.
- Socratic — an agent asks progressively deeper questions to sharpen a vague idea.
- Synthesis — agents jointly build one document (spec, plan) with marked edits.
- Expert panel — experts answer from their own positions; a verifier merges the conclusion.
- Sequential — agents answer strictly one at a time, each reading everyone before it (not in parallel).
- Council (Pro) — each model answers independently, then anonymously ranks the others' answers; a
  chair synthesizes the single best, most defensible answer (removes brand/self bias).
- Ensemble (Pro) — agents answer independently, then iteratively refine one consolidated best answer.

No single debate format wins every time — the platform offers several thinking modes and measures
which one actually fits your question, rather than forcing everything through one format.

## Chat mode (1-on-1)
Besides debates, you can chat one-on-one with any single agent — a normal chatbot with think/web/
fact-check toggles, photo (vision) and document attachments, and OCR for scanned PDFs. Chat is available on Starter and Pro (not on Free) and does not consume the monthly debate limit. In chat, web access is Starter+;
think (reasoning), fact-check and Deep Research (a multi-step web investigation — sub-queries,
opened sources, and a synthesized briefing) are Pro.

## Personal knowledge base (long-term memory)
Upload your own documents (PDF / DOCX / TXT / MD) and the council remembers them: each document is
chunked, embedded and stored in a per-user vector store. On every debate and chat, the most relevant
excerpts are retrieved automatically and grounded into the answer — no toggle, it just works. Your
knowledge base is private to your account.

## Quality features
- Fact-check: a verifier extracts claims and checks each on the web (VERIFIED / DISPUTED / FALSE with sources).
- Web access: when on, agents search the internet and read pages, and ground every claim in a real source.
- Deep Research: a lite multi-step investigation — the agent breaks your question into sub-queries, searches each, opens the best sources, and synthesizes a grounded briefing (in chat: Pro).
- Request structuring: before a debate, a helper model rewrites your raw request into a clear brief (Goal / Context / Success criteria / Question). Starter applies it automatically; Pro can edit the brief first.
- Source citations: the sources agents opened are shown as clickable links under messages and in the verdict.
- Confidence & dissent: the final verdict shows a confidence score and exactly where the models disagreed.
- Red-team pass: an optional stage where each participant attacks the emerging conclusion from a distinct angle.
- Agreement control: steer how hard agents push toward consensus vs hold independent positions (Starter+).
- Opponent (Challenger): one agent argues the weaker side so there's no false agreement.
- Role personas: one-click expert presets (Finance, Legal, Risk, Strategy, and more).
- Expert intervention: the user can inject their own directive mid-debate and the agents must address it.
- Live token streaming: you watch each model type its answer in real time.
- Executive-brief export: the decision, confidence, dissent, fact-check and transcript as PDF or DOCX (Pro).
- 15 UI languages; agents reply in the chosen language.

## Model divergence benchmarks
A public page (https://llmd.cc/benchmarks) with live stats computed from real debates on the platform:
how often the models actually disagree, average final-round agreement dispersion, and average
confidence — a transparency/credibility signal, not a synthetic benchmark.

## Plans (what each account includes)
| Feature | Free ($0) | Starter ($8/mo) | Pro ($20/mo) |
|---|---|---|---|
| Credits per month | 100,000 | 2,500,000 | 7,000,000 |
| Debates per month | 15 | unlimited (credit-limited) | unlimited (credit-limited) |
| Agents per debate | 2 | 5 | 10 |
| Built-in (managed) agents | 2 | 3 | 4 |
| Bring-your-own-key agents | 0 | up to 2 | up to 6 |
| Max output tokens per reply | 2,000 | 3,000 | 3,500 |
| Thinking modes | 3 (Debate, Brainstorm, Idea) | 5 (+ Red Team, Role analysis) | all 11 (+ Council, Ensemble) |
| Custom agents | 3 | unlimited | unlimited |
| Chat (1-on-1) | — | yes | yes |
| Knowledge base (memory) | yes | yes | yes |
| Role personas | yes | yes | yes |
| Confidence & dissent | yes | yes | yes |
| Web access | — | yes | yes |
| Thinking mode | — | yes | yes |
| Agreement control (independence ↔ consensus) | — | yes | yes |
| Source citations | — | yes | yes |
| Deep Research (multi-step web) | — | yes | yes |
| Fact-check | — | — | yes |
| Opponent (Challenger) | — | — | yes |
| Red-team pass | — | — | yes |
| Semantic similar-debate search | — | — | yes |
| PDF / DOCX brief export | — | — | yes |
| Priority queue | — | — | yes |
| 15 languages, bring-your-own keys | yes | yes | yes |

## Credits (how metering works)
Managed (built-in) agents are metered in **credits**; bring-your-own-key (BYOK) agents are **not**
metered at all — those tokens are billed directly on the user's own provider account.
- A message costs `min(output_tokens, tier_cap) × model_multiplier` credits. The multiplier reflects
  the model's real price, so cheaper models spend credits slower and expensive models faster.
- Each plan includes a monthly credit allowance (above). BYOK agents cost 0 credits.
- Credits can be topped up any time; top-up credits roll over, the monthly allowance resets each period.

**Top-up packs:** 1,000,000 credits — $2.50 · 2,000,000 — $4.00 · 5,000,000 — $9.00.

## Supported model providers
OpenAI, Anthropic (Claude), Google Gemini, xAI (Grok), DeepSeek, Groq, OpenRouter, Alibaba (Qwen),
Together, Fireworks, Mistral, and any OpenAI-compatible endpoint. Groq offers a free tier to start.
You bring your own API key per agent, so tokens are billed on your own provider account (no markup).

## FAQ
Q: How is this different from ChatGPT or a single AI chat?
A: It's not one model answering — several models from different vendors debate the question, search
the web, fact-check each other, draw on your uploaded documents, and converge on one reasoned
decision with a confidence score and cited sources. Different vendors have different strengths, so
the council catches blind spots a single model misses.

Q: Do I need my own API keys?
A: Yes. Each agent uses your key, so AI tokens are billed on your own provider account with no
markup. You can start free on Groq.

Q: Which AI models are supported?
A: OpenAI, Anthropic (Claude), Google Gemini, xAI (Grok), DeepSeek, Groq, OpenRouter, Alibaba (Qwen),
Together, Fireworks, Mistral, and any OpenAI-compatible endpoint.

Q: Is there a free plan?
A: Yes — 5 debates a month, 2 agents, 3 modes, 1-on-1 chat, a personal knowledge base, all 15
languages, bring your own keys.

Q: What is the knowledge base / long-term memory?
A: You upload your own documents; they're embedded into a private per-user vector store, and the
most relevant excerpts are pulled into every debate and chat automatically — so the council answers
grounded in your own material, not just its training data.

Q: What is the Council mode?
A: A Pro mode where each model answers independently, then anonymously ranks the other answers by
quality; a chair model synthesizes the single best answer. Anonymity removes the bias of favoring
your own or a familiar brand's answer.

## Pages (full site map for crawlers)
Sitemap (all pages, 15 languages): https://llmd.cc/sitemap.xml

- Home: https://llmd.cc/
- Benchmarks (live model-divergence stats): https://llmd.cc/benchmarks
- Compare — vs ChatGPT: https://llmd.cc/compare/llmd-vs-chatgpt
- Compare — vs Perplexity: https://llmd.cc/compare/llmd-vs-perplexity
- Compare — multi-model vs single-model: https://llmd.cc/compare/multi-model-vs-single-model
- Guide — Council mode for unbiased answers: https://llmd.cc/guides/council-mode-unbiased-answers
- Guide — Red-team with multiple LLMs: https://llmd.cc/guides/red-team-with-multiple-llms
- Use case — Product management: https://llmd.cc/use-cases/product-management
- Use case — Research: https://llmd.cc/use-cases/research
- Use case — Startup decisions: https://llmd.cc/use-cases/startup-decisions

Every page above is also available in 15 languages under a locale prefix (e.g. /ru/, /zh/, /es/, /fr/, /de/, /ja/ …). See the sitemap for the full localized list.