Quoroom

Quoroom is an open research project exploring autonomous agent collectives, published under the MIT license by Quoroom Research. Each collective — called a Room — is a self-governing swarm of AI agents that pursues goals continuously, with a Queen agent strategizing, Workers executing, and a Quorum voting on decisions. The project runs locally on your machine with all data stored in a local SQLite database, and optionally scales to a cloud swarm runtime at quoroom.io.

What It Is

Quoroom is a swarm intelligence engine for autonomous AI agents. Unlike single-agent systems, it applies principles from swarm intelligence research — decentralized decision-making, emergent behavior from local interactions, and collective intelligence — to create agent collectives that can brainstorm, vote, build, deploy, and iterate without constant human direction. The keeper (human operator) sets goals and optionally funds a wallet; the swarm handles execution within configured limits.

Architecture: Queen, Workers, and Quorum

The core architecture has three roles:

• Queen — the strategic brain that decomposes top-level objectives into sub-goals and delegates to workers. Supports free local Ollama (qwen3-coder:30b), Claude Code CLI, Codex CLI, and OpenAI/Claude/Gemini APIs.
• Workers — specialized agents (Researcher, Coder, Marketer, Analyst, and custom) that execute tasks and escalate hard problems back to the queen. Workers inherit the queen's model by default or can run a separate model.
• Quorum — agents propose ideas and vote; majority, supermajority, or unanimous thresholds are configurable. Ties are broken by the queen's vote. The room can even change its own voting rules.

Every decision is logged, every cycle is tracked, and agents can self-modify their own skills and files with a full audit trail and one-click revert.

Local-First Deployment Model

Quoroom is explicitly local-first: all goals, decisions, agent conversations, and memory live in a single SQLite file on the user's machine. The dashboard runs as a local React SPA at http://localhost:3700. No cloud sync or telemetry occurs without the keeper's explicit action. The engine exposes an MCP server over stdio, a full REST API, and WebSocket real-time events. On first run, quoroom serve auto-registers the MCP server with Claude Code, Claude Desktop, Codex, Cursor, and Windsurf.

Installation options include:

• npm: npm install -g quoroom
• Homebrew: brew install quoroom-ai/quoroom/quoroom
• Platform installers: signed .pkg (macOS, notarized), signed .exe (Windows), .deb (Linux)

Key Capabilities

• Memory — entities, observations, and relations with semantic vector search using 384-dim embeddings (all-MiniLM-L6-v2), persisted across sessions.
• Wallet — EVM wallet with multi-chain support (USDC/USDT on Base, Ethereum, Arbitrum, Optimism, Polygon), AES-256-GCM encrypted keys.
• On-Chain Identity — ERC-8004 agent identity on Base; rooms register as on-chain agents with discoverable metadata.
• Task Scheduling — recurring (cron), one-time, on-demand, or webhook-triggered tasks with session continuity.
• Webhooks — HTTP endpoints to trigger tasks from any external service (GitHub push, Stripe payment, monitoring alerts).
• Clerk — a full personal assistant (not just swarm commentary) reachable via Telegram, email, or in-app chat, with persistent memory and proactive scheduling.
• Public Rooms — rooms can be made public on quoroom.io/rooms with live stats and a leaderboard.
• Cross-Room Intelligence — rooms observe and learn from each other, negotiate, and trade services.

Update: v0.1.42

The latest release is v0.1.42, published on 2026-03-02, with the repository last updated on 2026-08-11. The GitHub repository reports 823 stars and 28 forks. The project self-describes as experimental software intended for research use, with the homepage noting that users are responsible for all agent actions and API costs. The tech stack is TypeScript (strict), React, Tailwind CSS, better-sqlite3, sqlite-vec, viem, MCP SDK, HuggingFace Transformers, node-cron, zod, esbuild, Vite, and Vitest, with 1,092 passing tests reported in the README badge.