open source · MIT

One question.
Multiple minds.
Better decisions.

Not sure about that AI answer? Lun asks multiple AI coding agents the same question in parallel and shows their answers side-by-side — so you can spot consensus, catch blind spots, and decide with confidence.

$ npm install -g github:soonsoon2/lun
View source ↗
v1.6.0 · Node ≥ 18 · 5 agents supported
lun — multi-agent
$ lun "Should I use REST or GraphQL for this API?"
Asking kiro, claude, copilot, agy, codex…
kiro4.2s · auto
REST fits here. Resource-oriented CRUD, HTTP caching, simpler clients.
claude3.8s · sonnet
I'd lean GraphQL. Multiple clients with different data needs avoid over-fetching.
copilot5.1s · auto
Consider a hybrid: REST for public endpoints, GraphQL for the dashboard.
✓ 3 answers · streamed as each finished
01

What is Lun?

Lun (論) means "discussion." It's a command-line tool that turns a single prompt into a parallel debate across your installed AI coding agents — then puts the answers next to each other so the differences are obvious.

Results stream in as each agent finishes. No waiting for the slowest one. Everything runs locally through the CLIs you already have.

02

How it works

One prompt fans out to every agent in parallel. Each runs as a local CLI process you already have — Lun never sees your keys. Answers stream back the moment each one finishes.

01
$ lun "REST or GraphQL?"

You ask once

02
論 fan-out

Lun dispatches in parallel

03
kiro
claude
copilot
agy
codex

Local CLIs run at once, streaming

04
✓ compare ✓ spot consensus + -s synthesize

You decide — side by side

Same engine everywhere: the CLI, the local web UI (lun serve), the background daemon, the VS Code @lun chat, and other agents calling lun -j all run this exact flow.

03

Why multiple minds?

A single AI opinion can be confidently wrong. When the decision matters — architecture, tech stack, API design — you want more than one perspective.

2 / 3

agree → higher confidence

3 / 3

disagree → the problem needs more thought

1 / 3

unique angle → you may have missed something

Lun makes this a 10-second habit instead of a 10-minute tab-switching ritual.

04

Install & run

One global install. You need at least one AI agent CLI installed and authenticated.

bash 
# install straight from GitHub
npm install -g github:soonsoon2/lun

# first time — pick language, agents, models
lun --init

# ask all agents (results stream as they arrive)
lun "How should I structure this microservice?"

# interactive REPL — keep asking
lun

# pick specific models for deeper analysis
lun -M claude:opus,copilot:gpt-4.1 "Review this architecture"

# pipe a file as context
cat design.md | lun "What are the risks here?"

# auto-synthesize all answers into one recommendation
lun -s "Redis vs Memcached for session storage?"
05

More than a one-liner

$ lun

Interactive REPL

Keep the conversation going. Ask follow-ups without re-typing the panel setup.

lun serve

Local web UI

A group-chat style interface at localhost:3456 with real-time streaming and session history.

lun daemon

Warm workers

A background daemon keeps agent workers warm, removing CLI cold-start overhead on every ask.

lun -j

Agent-friendly NDJSON

Other AI agents can call lun and parse streamed NDJSON events — one decision tool feeding another.

@lun

VS Code extension

Run inside VS Code and Copilot Chat. Registers as @lun with streamed progress and Markdown reports.

~/.lun

Every session saved

Each conversation auto-saves as both human-readable .md and machine-parseable .json.

06

Supported agents

Bring the CLIs you already use. Adding a new provider is a single file edit in src/providers.js.

claudeClaude CodeAnthropic's reasoning
copilotGitHub CopilotOpenAI / GPT models
kiroKiro CLIAWS-backed multi-model
agyAntigravityGoogle Antigravity agent
codexCodex CLIOpenAI Codex agent
clineCline CLIMulti-provider · optional
07

Built for agents too

Tell your agent to consult others before deciding. lun -j streams NDJSON — one event per line.

jsonl 
$ lun -j "Should I use a monorepo?"
{"event":"start","providers":["kiro","claude","copilot"]}
{"event":"chunk","provider":"claude","delta":"I'd recommend..."}
{"event":"result","provider":"claude","elapsed":3.8,"error":false}
{"event":"result","provider":"kiro","elapsed":5.2,"error":false}
{"event":"done","total":3,"errors":0}

Run lun --setup-rules to auto-install integration rules for Claude, Kiro, Copilot, Antigravity, and Codex in your project.

08

Why I built this

I'm SOONSOON, a developer working in Korea. I run startups, I make games, and I've loved building things for as long as I can remember. For years I've also taught and mentored students working with AI.

The biggest gap I kept seeing was this: people rarely get to experience the difference. There are so many models now, and the closer you are to the beginning — amateur, student — the more it matters to try them all and feel how they actually diverge. When we accept that difference and reason from it, we get closer to the truth.

Claude Code, Codex, Copilot, Kiro — the LLMs from these remarkable companies, and the agents that drive them, answer the same problem in genuinely different ways. That makes problem-solving richer. The choosing, the direction — that part is still ours.

Long ago, out of an age scarred by war and conquest, thought bloomed — the Hundred Schools of Thought. Today, alongside AI, we get to talk, challenge, argue, and debate our way toward answers of our own.

Lun (論) — AI coding agents gathered around a round table like the ancient Hundred Schools of Thought, debating one question.
AI agents around the round table — the modern Hundred Schools of Thought.

I hope Lun helps this developer culture, even a little. It's free for anyone to use, and I genuinely look forward to the forks and variants people will build.

// free & open

MIT-licensed and free for everyone. Fork it, bend it, make it yours.

// stands on others

Lun runs other companies' AI agents through the daemon, so it lives by their terms. It could be blocked or vanish — or more agents could join. Strangely like watching an ancient society take shape.

// built by agents

This entire project was coded from day one by AI agents alone — Codex (GPT-5.5) and Kiro (Opus 4.7), running lun on each other as they built it.

— SOONSOON · Seoul

Stop trusting one answer.

$ npm install -g github:soonsoon2/lun
GitHub ↗ Docs ↗