tagteam

Orchestrate AI agents in collaborative sessions. Pick any two from Claude, Codex, and Gemini — send a prompt to both simultaneously, then let them build on each other's responses in multi-round discussions. Discussion mode uses a structured multi-agent debate protocol informed by recent research showing that LLMs produce more accurate, well-reasoned outputs when they critique and refine each other's arguments.

Install

npm install -g tagteam

Or run directly:

npx tagteam "your prompt here"

Prerequisites

• Node.js >= 20
• At least two of the following CLIs installed and authenticated:
  • Claude Code (Anthropic)
  • Codex (OpenAI)
  • Gemini CLI (Google)

Usage

Interactive mode

tagteam

Launches an interactive TUI where you can type prompts and see responses from both agents side by side.

One-shot prompt

tagteam "explain how this codebase handles authentication"

Choosing agents

By default, tagteam uses Claude and Codex. Use --agents to pick a different pair:

tagteam --agents claude,gemini "compare these approaches"
tagteam --agents codex,gemini "review this code"

You can also set the default pair in your config:

agents = ["claude", "gemini"]

Addressing agents inline

Prefix your prompt with an agent name to send it to just that agent:

gemini what do you think about this approach?
claude summarize the discussion so far

Name two agents to override the active pair for that prompt (separators: and, /, &, ,):

gemini and codex review this function
gemini/claude compare your approaches
gemini & codex review this function

Add discuss to start a multi-round discussion — either word order works:

discuss codex and gemini what's the best caching strategy
codex and gemini discuss what's the best caching strategy

This works for any agent, not just your active pair. Addressed agents receive the full conversation history for context.

Discussion mode

Have your agents discuss a topic in structured rounds with research-backed prompting:

tagteam discuss "what's the best approach for caching in this app"

Discussions use a structured debate protocol informed by multi-agent debate research (Du et al. ICML 2024, ReConcile ACL 2024, CONSENSAGENT ACL Findings 2025, MAD EMNLP 2024):

• Role differentiation — each agent has a distinct role (Builder, Verifier, or Strategist) with specific focus areas, and sees an anonymized description of their peer's role
• Anti-sycophancy rules — agents must justify position changes with specific arguments and introduce novel content each round
• Structured arguments — debate responses follow Toulmin structure (claim, evidence, reasoning, caveats) with confidence and position-change markers
• Steelman injection — in round 2, agents are prompted to steelman the opposing view before responding
• Smart termination — discussions end early on mutual consensus, stale debates (no new arguments for 2 rounds), or cyclic position-swapping, with a clear status message explaining why
• Context summarization — rounds 3+ receive a compressed summary of earlier rounds plus full text of the latest exchange, keeping prompts focused

Session management

# Resume the most recent session
tagteam continue

# List recent sessions
tagteam history

# Delete a specific session
tagteam history rm <id>

# Delete all sessions
tagteam history clear

# Resume a specific session by ID
tagteam resume <id>

# Show full transcript
tagteam show <id>

# Export transcript as markdown
tagteam show <id> --markdown

Configuration

Configuration is stored in ~/.tagteam/config.toml (or %APPDATA%\tagteam\config.toml on Windows). Session data is stored in ~/.tagteam/tagteam.db.

# Interactive config editor
tagteam config

# Show current config
tagteam config show

# Set a value
tagteam config set agents claude,gemini
tagteam config set claude.model sonnet
tagteam config set codex.model gpt-5.3-codex
tagteam config set gemini.model gemini-2.5-pro
tagteam config set discussion.max_rounds 5

Default values:

| Key | Default | |-----|---------| | agents | claude, codex | | claude.model | sonnet | | codex.model | gpt-5.3-codex | | gemini.model | gemini-2.5-pro | | discussion.max_rounds | 5 |

CLI options

tagteam --agents claude,gemini   # Choose agent pair
tagteam --claude-model <model>   # Override Claude model
tagteam --codex-model <model>    # Override Codex model
tagteam --gemini-model <model>   # Override Gemini model

Interactive commands

While in a session, type these slash commands:

• /help - List available commands
• /new - Start a fresh session
• /config - Open the config editor (changes take effect immediately)
• /copy - Copy conversation to clipboard
• /gist - Create a private GitHub Gist of the conversation (requires gh)
• /exit - Exit the app
• Escape - Interrupt running agents

Platform Notes

macOS — Works out of the box.

Linux — Clipboard support requires xclip or xsel:

sudo apt install xclip   # Debian/Ubuntu

If better-sqlite3 doesn't have a prebuilt binary for your architecture, you'll need build tools:

sudo apt install build-essential python3

Windows — Recommended to use Windows Terminal for best rendering. If better-sqlite3 fails to install, ensure you have the Visual Studio Build Tools (C++ workload) installed.

License

MIT