opendisc

English | 简体中文

An efficient CLI tool that enables multi-agent debate to produce better answers (currently supports Codex and Claude Code). Examples

Features

• Multi-agent collaborative discussion with flexible order, including same-agent debate (for example claude vs codex, codex vs codex, claude vs claude).
• Multi-round iteration with a default of 2 rounds (configurable).
• Built-in history management: start discussions, list records, show full transcript, or show final answer only.

Installation

1) Install and authenticate required CLIs

npm install -g @anthropic-ai/claude-code
npm install -g @openai/codex

Make sure both claude and codex are authenticated and runnable in your shell.

2) Install opendisc

npm install -g @jiangz222/opendisc
opendisc --help

Usage

Basic

opendisc "What feature modules are needed to build an MVP version of an AI assistant?"

Config file

~/.opendisc/config.json

Example:

{
  "maxRounds": 3,
  "agentOrder": ["claude", "codex"]
}

Supported fields:

• maxRounds: positive integer
• saveDir: directory for discussion records (optional)
• agentOrder: any two-agent pair from claude/codex (for example ["claude","codex"], ["codex","codex"])
  • Recommended: ["claude","codex"]
• outputLanguage: output language (optional), supports zh-CN or en-US

CLI flags override config file values (such as --max-rounds, --save-dir, --agent-order, --output-language).
Priority: CLI flags > config file > automatic detection (from topic text).

Common commands

# Set max rounds
opendisc "What feature modules are needed to build an MVP version of an AI assistant?" --max-rounds 3

# Set model order
opendisc "Design an auth strategy" --agent-order codex,claude

# Same-agent debate
opendisc "Review this migration plan" --agent-order codex,codex

# Set output language (higher priority than auto detection)
opendisc "Design a database indexing strategy" --output-language en-US

# List discussion records (default dir: ~/.opendisc/discussions)
opendisc list

# Show full record
opendisc load 2026-02-28T10-00-00-000Z_topic.json

# Show final answer only
opendisc show 2026-02-28T10-00-00-000Z_topic.json

show output always contains two sections:

1. 1 Final Conclusion: a direct, actionable, single final answer to the original question.
2. 2 Key Disagreements (for reference): key disagreements only; explicitly marks none when there are no key disagreements.

Examples

For example: Through multi-round discussion, Codex lowered the priority of RAG in a personal AI assistant MVP.

Round 1

Round 2

Final Conclusion

Dependencies

• Core dependencies:
  • @anthropic-ai/claude-code for Claude CLI/SDK integration
  • @modelcontextprotocol/sdk for Codex MCP integration
  • commander for CLI argument parsing
  • chalk for terminal styling
• Implementation reference:
  • src/vendor/claude/sdk/* and src/vendor/codex/* are adapted from ideas used in slopus/happy, simplified for opendisc.

License

MIT