EXPERIMENTAL RESEARCH PROJECT

What if multiple LLMs debated your code before writing it?

A research experiment: multiple LLMs debate your task in structured rounds before implementing. Uses your existing Claude Code or Gemini CLI. Expect bugs. This is not production software.

$ curl -fsSL https://raw.githubusercontent.com/adndvlp/conclave/main/install.sh | bash
How it works GitHub

How Conclave Works

You ask
"Build auth system"
Team debates
Multi-round deliberation
Converges
By merit & consensus
Implements
Production code
Conclave 1 (2 models)
━━━━━━━━━━━━━━━━━━━━━━━━━━━
Team debating · Round 1/3
DeepSeek V4 Flash: LEAD — Auth system architecture
GLM-5.1: SUPPORT:DeepSeek — Backend expert
━━━━━━━━━━━━━━━━━━━━━━━━━━━
Converged — DeepSeek V4 Flash implements

Standing on OpenCode's shoulders

OpenCode
/connect
/models
/agents
git
tools
LSP
providers
TUI
+
Team Layer
/team
debate.ts
prompts.ts
team.ts
processor
signals
=
Conclave

OpenCode already had providers, agents, git, tools, and a polished TUI. Conclave adds one thing: a team debate engine. ~12 files modified. The rest is OpenCode doing what it does best.

Does Team Debate Outperform Solo?

Democratic Consensus

No single model decides. Structured rounds with LEAD, SUPPORT, ALIGN, BUILD, CHALLENGE signals. Winner chosen by endorsement score.

Blind Spot Detection

One model catches what another misses. Architecture flaws, edge cases, and security issues caught during deliberation.

Live Reasoning

Watch each model think in real-time. Click to expand reasoning blocks. Round summaries always visible.

CLI Adapters

Connect Claude Code, Gemini CLI, or Codex as team members. No API keys needed — uses your existing CLI auth. Perfect for those with subscriptions but no API access.

Any Provider

OpenAI, Anthropic, DeepSeek, Google, NVIDIA, Groq, Ollama. Mix API and CLI models in the same team.

Persistent Teams

Multiple named teams with different model combos. Survive restarts. Switch with a single command.

Breaking Teams

LLMs autonomously split into sub-teams for complex tasks. Frontend and backend in parallel with cross-team communication.

This is an experiment. Not a product.

Any
Team Size
Rounds
MIT
License
All
Providers

Start Your Conclave

$ curl -fsSL https://raw.githubusercontent.com/adndvlp/conclave/main/install.sh | bash

What to Expect

Context windows are real

Each model has its own context limit. Conclave auto-truncates the debate thread per model — large-context models see everything, small-context models see signal summaries. No crashes, no compaction failures. Pair a 1M-token model (DeepSeek) with a 128K model (Gemini Flash) for optimal depth vs speed.

Latency is additive

3 models debating 3 rounds = 9 API calls per message. Responses take longer than a single LLM. The tradeoff: better output quality for more complex tasks. For simple questions, switch to single model mode.

Cost mirrors team size

Each team member makes independent API calls. A 3-model team costs 3× more per message. Free tiers (Gemini CLI, Ollama local models) offset this. Choose your team based on task complexity, not just model quality.

This is v1

Conclave is weeks old. The debate engine works, teams persist, providers connect. But expect rough edges. Context optimization, live streaming, and Breaking Teams are on the roadmap. Feedback and contributions welcome.