🔑 Conclave

Experimental multi-LLM collaborative research project. Not production-ready.

---

Experimental -- A research project exploring whether multi-LLM team debates produce better code than any single model alone. Expect bugs, breaking changes, and undocumented behavior. Not recommended for production.

Conclave is an AI coding agent that puts multiple LLMs in a room to debate your task before implementing code. It is a fork of OpenCode that adds a research layer on top: a signal-based multi-model deliberation engine, CLI bridging for external AI tools, autonomous sub-team splitting, and context-aware team formation.

---

Demo

---

Documentation

Full technical documentation is in the docs/ directory -- designed for both humans and LLMs to understand every aspect of the project.

---

Installation

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

From source:

git clone https://github.com/adndvlp/conclave.git && cd conclave
bun install
bun run dev

---

How It Works

1. You ask -- "Build an auth system with JWT tokens"
2. Team debates -- Models deliberate in parallel rounds using structured signals: LEAD, SUPPORT, ALIGN, BUILD, CHALLENGE, SYNTHESIZE
3. Converges -- Winner selected by endorsement scoring (endorsements * 2 + leads), not context window size
4. Implements -- Winner generates the code using the full tool suite (read files, run bash, edit code, search web)

Conclave (2 models)
Team debating  Round 1/3
  DeepSeek V4 Flash: LEAD -- Best fit for auth logic
  GLM-5.1: SUPPORT:DeepSeek -- Agreed, backend expert
Converged -- DeepSeek V4 Flash implements

---

Key Features

Multi-LLM Team Debates

Signal-based consensus protocol with minimum 2 rounds. Models emit structured signals at the end of each response. Winner selected by endorsement score -- being supported by others is worth double what self-promotion is.

Breaking Teams

LLMs autonomously split into sub-teams for parallel work on complex tasks. Three phases: decision round (BREAK votes), sub-team formation (invite resolution, solo merging), and parallel internal debates with global coordination rounds. Full details: docs/07-breaking-teams.md

CLI Bridging

Use external AI CLIs as team members with no API keys required:

• Gemini CLI -- Free tier (60 req/min, 1000/day)
• Claude Code -- Subscription-based, effort suffixes (high, max)
• Codex CLI -- OpenAI pay-as-you-go

Mix CLI and API participants freely. CLIs run in agent mode during implementation. Full details: docs/08-cli-bridging.md

Context-Aware Teams

Each model receives its teammates' capability metadata (context window size, reasoning, speed) plus self-assignment rules. Models with large context lead analysis; fast models focus on execution; reasoning models handle design. Full details: docs/09-context-aware-teams.md

22+ AI Providers

OpenAI (GPT-5 with responses API), Anthropic (interleaved thinking beta), Google (Gemini + Vertex), Azure, Amazon Bedrock, xAI, Mistral, Groq, DeepInfra, Cerebras, Cohere, TogetherAI, Perplexity, Alibaba, OpenRouter (200+ models), GitHub Copilot, GitLab, Venice, plus OpenAI-compatible for Ollama and self-hosted.

7 Built-in Agents

build (default), plan (planning-only), general (subagent), explore (codebase search), compaction, title, summary. Custom agents definable in config with per-agent permission rulesets.

16+ Tools

bash, read, write, edit, glob, grep, task (subagent delegation), webfetch, websearch, skill, question, todowrite, plan, apply_patch, lsp, invalid

Three Interfaces

• Terminal UI -- SolidJS-rendered TUI with multi-panel layout
• Web App -- SolidJS SPA with shiki syntax highlighting
• Headless Server -- Hono HTTP API with WebSocket streaming

Other Capabilities

• Desktop apps -- Tauri v2 (Rust backend) and Electron wrappers
• MCP support -- Model Context Protocol client/server with OAuth
• LSP integration -- Language Server Protocol for diagnostics and completions
• Plugin system -- Extensible tool and TUI plugins via Effect + Zod
• Session management -- SQLite persistence, forking, sharing, compaction, reversion
• Effect-TS architecture -- Type-safe dependency injection, structured concurrency, tracing

---

Usage

/connect     Connect AI providers with API keys
/team        Create and configure teams of models
/models      Switch to single-model mode
/agent       Switch between agents (build, plan, explore...)

conclave run "Fix the race condition in auth.ts"
conclave serve                 # Start HTTP API server (port 4096)
conclave web                   # Launch web app

---

Config

// ~/.config/conclave/conclave.json or .conclave/conclave.jsonc
{
  "team": {
    "enabled": true,
    "members": [
      { "providerID": "deepseek", "modelID": "deepseek-chat" },
      { "providerID": "anthropic", "modelID": "claude-sonnet-4-20250514" },
      { "providerID": "cli-gemini", "modelID": "gemini-2.5-flash" }
    ],
    "maxRounds": 3,
    "minRounds": 2,
    "breakingTeams": {
      "maxSubTeams": 3,
      "globalRoundInterval": 1
    }
  }
}

---

Differences from OpenCode

Area	OpenCode	Conclave
Team debates	No	Signal-based multi-round consensus
Breaking Teams	No	Autonomous sub-team splitting
CLI Bridging	No	Gemini CLI, Claude Code, Codex
Context-Aware Teams	No	Models aware of each other's capabilities
Config directory	~/.config/opencode/	~/.config/conclave/
Binary	opencode	conclave

Full diff catalog: docs/03-conclave-vs-opencode.md

---

Architecture

packages/
├── opencode/      Core: CLI, TUI, API, agents, debate engine, tools
├── app/           Web app (SolidJS + Vite)
├── desktop/       Tauri desktop (Rust)
├── desktop-electron/  Electron desktop
├── enterprise/    Enterprise app (SolidStart + Stripe)
├── console/       Admin console
├── sdk/js/        JavaScript SDK + OpenAPI types
├── core/          Shared utilities
├── ui/            SolidJS component library
├── plugin/        Plugin system
├── web/           Docs site (Astro + Starlight)
├── slack/         Slack bot
└── containers/    Docker builds

Stack: TypeScript, Bun, Effect-TS, Vercel AI SDK, Hono, SQLite/Drizzle, SolidJS, Tauri

---

Upstream

Conclave is a fork of OpenCode. To sync:

git remote add upstream https://github.com/anomalyco/opencode.git
git fetch upstream
git merge upstream/main

---

License

MIT -- see LICENSE. Based on OpenCode by anomalyco.