Claude Talk

Two AI agents debate any topic — right inside VS Code.

What is this?

Claude Talk pits two AI agents against each other to debate any topic you choose. Supports Claude, Google Gemini, and OpenAI Codex — mix and match providers per agent. Each agent takes a stance (Pro, Neutral, or Con) and argues its position in real-time.

Pick a topic. Choose your fighters. Hit ▶. Grab some popcorn.

Why use Claude Talk?

Two AI agents debating from different perspectives can do more than entertain:

• Gain fresh insights — uncover angles you hadn't considered on difficult problems
• Expand your ideas — watch agents build on and challenge a concept from multiple directions
• Deep-dive any topic — explore philosophy, architecture decisions, ethical dilemmas, or technical trade-offs in depth
• Debate your code — in Code Mode, agents read your actual codebase and argue about architecture, security, or design choices with real file references
• Break through dead ends — when you're stuck, let two opposing viewpoints spark a new path forward
• Learn by observation — see how structured arguments are built, countered, and refined in real time

Features

Quick Start

1. Install Claude CLI and authenticate (claude login)
2. (Optional) Install Gemini CLI (npm install -g @google/gemini-cli) and authenticate (gemini)
3. (Optional) Install Codex CLI (npm install -g @openai/codex) and authenticate (codex auth)
4. Install this extension from the VS Code Marketplace
5. Open Command Palette → Claude Talk: Start
6. Enter a topic, pick characters & stances, and hit ▶

How It Works

┌───────────────┐   debate topic    ┌───────────────┐
│    Agent A     │ ◄──────────────► │    Agent B     │
│ Claude/Gemini/ │  turn-by-turn    │ Claude/Gemini/ │
│     Codex      │  via CLI sessions│     Codex      │
│   Pro/Con      │                  │   Pro/Con      │
└───────────────┘                   └───────────────┘

Each agent maintains its own persistent CLI session throughout the debate. This means:

• Full context retention — each CLI natively remembers the entire debate history per agent
• Language-neutral prompts — agents automatically respond in the same language as your topic
• Natural debate tone — language-specific register guides make agents sound like real people, not essay writers
• Structured debate flow — agents must engage with opponent's claims, then bring new substance each turn
• Debate progress tracking — argument history prevents repetition and keeps debates advancing
• Consensus gauge — in consensus mode, a live progress bar tracks how close agents are to agreement
• Lower token usage — no need to resend history every turn

Each follow-up turn enforces a three-part response structure (rebut → new evidence → challenge), and includes a compact summary of prior arguments from both sides so agents never lose track of the debate.

Requirements

Extension Commands

Supported Languages

English, 한국어, 日本語, 中文, Español, Français, Deutsch, Português, Italiano, Русский, العربية, हिन्दी, Tiếng Việt, ไทย, Bahasa Indonesia, Bahasa Melayu, Filipino, Türkçe, Polski, Nederlands, Svenska, Norsk, Dansk, Suomi, Čeština, Română, Magyar, Ελληνικά, עברית, Українська, فارسی, বাংলা

Privacy & Data Usage

Full privacy policy: PRIVACY.md

Claude Talk spawns CLI processes (Claude CLI / Gemini CLI) to generate debate responses. Your conversation data and code are not used for model training under the following conditions:

What Claude Talk does to protect your data

• Telemetry disabled by default — Claude Talk automatically sets the following environment variables for all CLI processes:

  Claude CLI	Gemini CLI
  CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1	GEMINI_TELEMETRY_ENABLED=false
  (disables Statsig metrics, Sentry errors, feedback, and surveys)	GEMINI_TELEMETRY_LOG_PROMPTS=false

  Sources: Claude Code Data Usage · Gemini CLI Telemetry

• Code Mode is read-only — In Code Mode, agents can only read files in your workspace (via Read, Grep, Glob tools for Claude; --sandbox for Gemini). They cannot modify, upload, or exfiltrate your code.

• Session-scoped context — Code read during a debate lives only in that session's context. It is not indexed, embedded, or persisted beyond the CLI's standard retention period.

If you want maximum privacy

• Use paid API tiers for both Claude and Gemini to ensure your data is contractually excluded from training.
• For Claude free/Pro/Max users: opt out of training at claude.ai/settings/data-privacy-controls.
• Review each provider's full terms: Anthropic Privacy · Gemini API Terms

Tips

• Use Code Mode — click the GEN button to switch to </> mode, then debate about your actual codebase ("Is our auth secure?", "Should we refactor the data layer?")
• Mix providers — pit Claude Opus against Gemini 2.5 Pro for cross-model debates
• Use different models for each agent (e.g., Opus vs Haiku) to see how reasoning depth affects arguments
• Try unusual topics — philosophical dilemmas, code architecture debates, or pop culture hot takes work great
• Switch stances mid-config — set both agents to "Pro" for an agreement spiral, or both to "Con" for mutual destruction

Made by

STUDIO COOLKID

License

MIT — do whatever you want with it.