↓ Skip to main content
Yonk-Labs
Yonk-Labs
  1. Projects/

abe

·3 mins
Rust Agents Llm Debate Code-Review Mcp
Table of Contents

abe (named for Lincoln, one of history’s great debaters) puts more than one model on a question instead of trusting a single answer. It broadcasts a prompt to several LLMs, runs them through a configurable number of critique rounds, and returns a synthesized final answer along with a report on where the models agreed and where they didn’t.

Why abe?
#

  • A second opinion, automatically. A second model reads the first’s answer and disagrees where it would — catching the confident-but-wrong case a single pass misses.
  • Works with zero cloud config. Models can be HTTP providers (OpenAI, Anthropic, any OpenAI-compatible endpoint) or local CLIs (codex, claude, opencode), mixed freely in the same debate. Pair two CLI providers and there’s no API key to manage at all.
  • Grounded in your material. Attach a design doc, a README, a PR diff — the debate argues over your actual document, not a generic prompt.
  • Personas give the panel real disagreement. Assign one model a security lens and another an SRE lens instead of two models politely agreeing with each other.
  • Three surfaces. A CLI, a stdio MCP server for Claude Code / Codex / opencode, and a small web UI with a JSON API.

What it’s for
#

  • Second-guessing a decision — abe validate runs one reviewer against a statement you’re about to act on.
  • Full debate on a design question — abe debate runs a multi-round argument across every configured model and returns a synthesized verdict.
  • In-editor validation from an agent — as an MCP server, Claude Code or opencode can call debate/validate directly mid-session instead of you copy-pasting prompts elsewhere.
  • Reviewing a diff or spec with distinct lenses — attach the file, assign personas per model, and get a panel argument instead of one model’s take.

How it works
#

Abe broadcasts the prompt to every configured model concurrently, then for each critique round shows each model the others’ latest (anonymized) answers and asks it to revise. A decision protocol — synthesis (a chairman model merges everything), judge (a judge model scores and picks one verbatim), or majority (deterministic clustering, no extra model call) — produces the final answer. Raw per-model answers are always preserved in the result.

Quick start
#

curl -fsSL https://raw.githubusercontent.com/yonk-labs/abe/main/install.sh | sh
abe init                                             # interactive model setup
abe debate "Is Postgres a good default database?"
abe validate --reviewer codex "We should rewrite this service in Rust."

Links#

  • GitHub Repository
  • Related: bob (uses abe as its judge)
  • Related: lede (abe’s --lede flag compresses oversized --files context without an extra LLM call)
  • License: MIT