versus-llm 1.0.1

Two LLMs enter. One verdict leaves.

versus-llm

Two LLMs enter. One verdict leaves.

A Python CLI/TUI where two LLMs debate a question across N rounds — Agent A argues for, Agent B challenges — and then a Judge delivers a final verdict. Powered by OpenRouter, with a Rich terminal UI.

# Classic CLI mode (pass a question)
versus "Is PostgreSQL better than MongoDB?" --file schema.sql --rounds 2

# Interactive TUI mode (no question)
versus

Install

pip install versus-llm

For local development from this repository:

pip install -e .

Setup

You need your own OpenRouter API key. Get one here: https://openrouter.ai/keys

versus setup

versus login does the same thing. The key is saved locally in your user config directory, so you do not need to create a .env file manually.

OPENROUTER_API_KEY is still supported as an environment variable, and a .env file in the current working directory is still supported.

Usage

versus "React or Vue?"
versus "Is Rust worth learning in 2025?"
versus "Which database fits this schema?" --file schema.sql
versus "Best state manager for React?" --rounds 3 --models google/gemma-4-31b-it:free meta-llama/llama-3.3-70b-instruct:free
versus "Is PostgreSQL better than MongoDB?" --output debate.md   # save transcript
versus --version

Arguments

Arg	Description	Default
question	positional; omit it to launch the interactive TUI	—
--file	path to a file for additional context (max 200KB)	none
--rounds	number of debate rounds	2
--models	two model slugs (Agent A then Agent B)	google/gemma-4-31b-it:free and meta-llama/llama-3.3-70b-instruct:free
--output	save the full debate transcript as Markdown to a path	none
--version	print versus-llm 1.0.1 and exit	—

Commands

Command	Description
versus setup	save your OpenRouter API key to local config
versus login	same as versus setup

Note on free models: the defaults are free-tier models and can occasionally return a transient 429 ("rate-limited upstream") when providers are busy. versus automatically retries up to 3 times with backoff (honoring the server's Retry-After), so most hiccups recover on their own. If it still fails, add your own provider key in your OpenRouter settings, or swap in other models via --models, e.g.:

versus "your question" --models google/gemma-2-9b-it:free meta-llama/llama-3.3-70b-instruct:free

Interactive TUI

Run versus with no question to launch a full-screen terminal app:

1. Input screen — type your question, an optional context file, and pick the number of rounds (1–5).
2. Debate screen — Agent A (left, 🟠) and Agent B (right, 🔵) stream their arguments in real time, then the Judge's verdict (⚖️) appears at the bottom.

Keybindings

Key	Action
S	Save the transcript to debate.md
r	Restart with a new question
q	Quit
Ctrl+C	Quit (press twice within 3s to confirm)

How it works

1. Agent A (🟠) makes the strongest case for / proposes the best solution.
2. Agent B (🔵) challenges Agent A, finds flaws, proposes alternatives.
3. Steps 1–2 repeat for --rounds rounds.
4. The Judge (⚖️) reads the full transcript and delivers a decisive verdict.

When --file is supplied, its contents are prepended as context to the question for both agents.