DebateTalk

npm version npm downloads License: MIT

Official MCP server and CLI for DebateTalk — run structured multi-model AI debates from your AI assistant or terminal.

DebateTalk makes multiple AI models argue a question independently, challenge each other's reasoning, and converge on a structured synthesis: Strong Ground, Fault Lines, Blind Spots, and Your Call.

---

Demo

DebateTalk demo

---

Features

• MCP server — connect Claude Desktop, Cursor, or any MCP-compatible client to DebateTalk
• CLI — run debates and check model status from the terminal
• Streaming progress — real-time round-by-round updates via MCP logging notifications
• Cost & share links — every debate shows cost and a shareable URL
• 5 tools: run_debate, get_model_status, recommend_models, estimate_cost, get_history

Try without an API key — dt models works instantly, no signup needed. See which models are online and their live latency. For debates, recommendations, and history you need an API key — free accounts include one at console.debatetalk.ai/api-keys.

---

Install

npm install -g @debatetalk/mcp

This gives you the dt CLI and the MCP server.

Set your API key:

export DEBATETALK_API_KEY=dt_your_key_here

Get a key at console.debatetalk.ai/api-keys. Free accounts include 1 API key — sign up at console.debatetalk.ai.

---

CLI

Run a debate:

dt debate "Should we adopt microservices?"

Example output:

- Starting debate…
✔ Debate complete

"Should we adopt microservices?"
Cost: $0.0842

━━━ STRONG GROUND ━━━
1. Microservices suit orgs with 50+ engineers and distinct team boundaries
2. Monoliths are simpler to operate at small scale
3. Migration costs are consistently underestimated

━━━ FAULT LINES ━━━
1. If teams need independent deploy cycles, then microservices — otherwise monolith
2. If operational maturity is low, then monolith — otherwise microservices are viable
3. If data consistency is critical, then monolith avoids distributed transaction pain

━━━ BLIND SPOTS ━━━
1. Cost of distributed tracing and observability tooling
2. Team cognitive load increases with service count
3. Shared database patterns can bridge the gap

━━━ YOUR CALL ━━━
1. Do you have dedicated platform engineering?
2. Are teams already blocked on shared deploy cycles?
3. Can you afford 6+ months of migration investment?

🔗 https://console.debatetalk.ai/share/a1b2c3d4-5678-90ab-cdef-1234567890ab

Check which models are online:

dt models

Get a recommended model panel for your question:

dt recommend "Is Rust worth learning in 2026?"

Estimate cost before running:

dt cost "Should we raise our Series A now?"

View past debates:

dt history
dt history --limit 5

---

MCP Server (Claude Desktop, Cursor, Cline, Goose, and any MCP client)

Add to your MCP client config:

{
  "mcpServers": {
    "dt": {
      "command": "npx",
      "args": ["-y", "@debatetalk/mcp"],
      "env": {
        "DEBATETALK_API_KEY": "dt_your_key_here"
      }
    }
  }
}

Config file locations:

• Claude Desktop (Mac): ~/Library/Application Support/Claude/claude_desktop_config.json
• Claude Desktop (Windows): %APPDATA%\Claude\claude_desktop_config.json
• Claude Code: ~/.claude/settings.json (under mcpServers)
• Cursor: .cursor/mcp.json in your project root
• Windsurf: ~/.codeium/windsurf/mcp_config.json
• Cline / Roo Code: MCP settings panel in VS Code extension
• Goose: ~/.config/goose/config.yaml (under extensions)
• Other clients: refer to your client's MCP documentation

Ask your AI assistant to run a debate

MCP clients read the tool description to decide when to call it — no exact phrasing required. Any of these work:

"debate whether we should rewrite our backend in Go" "use DT — should we raise our Series A now?" "multi-model this: is Rust worth learning in 2026?" "stress-test this architecture decision" "get a second opinion on moving to microservices"

Claude will also invoke it proactively for high-stakes decisions where a single AI answer is insufficient.

Example — Cursor:

Ask Cursor's AI chat: "use the run_debate tool — should we switch from REST to GraphQL?"

Cursor calls the MCP tool, streams progress, and returns the full synthesis inline in your chat.

Example — Cline (VS Code):

In Cline's agent chat: "debate: is our current auth middleware secure enough for SOC 2?"

Cline detects the run_debate tool from the MCP server and runs the debate. Results appear in the Cline output panel.

---

Claude Code — plugin marketplace

1. Add the DebateTalk marketplace:

/plugin marketplace add DebateTalk-AI/mcp

2. Install the plugin:

/plugin install debatetalk@debatetalk-mcp

3. Set your API key:

Add to ~/.claude/settings.json:

{
  "pluginConfigs": {
    "debatetalk@debatetalk-mcp": {
      "options": {
        "api_key": "dt_your_key_here"
      }
    }
  }
}

Then run /reload-plugins — the five DebateTalk tools are immediately available in your session.

---

MCP Tools Reference

Tool	Auth required	Description
run_debate	Yes	Run a structured multi-model debate (streaming)
get_model_status	No	Real-time health and latency for all models
recommend_models	Yes	Get the best model panel for your question
estimate_cost	Yes	Estimate credit cost before running
get_history	Yes	List your past debates

run_debate

question     string   required   The question or topic to debate
models       array    optional   Specific model IDs to use (omit for smart routing)
rounds       number   optional   Number of deliberation rounds (default: 2)

get_model_status

No parameters. Returns live health, latency, and uptime per model.

recommend_models

question     string   required   The question — routing picks the strongest panel

estimate_cost

question     string   required
models       array    optional
rounds       number   optional

get_history

limit        number   optional   Number of debates to return (default: 20, max: 100)

---

Contributing

See CONTRIBUTING.md. Issues and PRs welcome.

---

License

MIT — see LICENSE.