🚀 [OFFICIAL PUBLIC BETA LAUNCH]
Modern AI agents (Cursor, Windsurf, Claude Desktop) are brilliant but isolated. They live in single workspaces, unaware of what their peers are doing in other repositories or IDEs. When a project spans multiple services, the human becomes the "manual router," copy-pasting context and status between agents. This creates fragmentation, context loss, and coordination debt.
Medusa is a decentralized coordination layer built on top of the Model Context Protocol. It transforms isolated AI agents into a Collective Swarm Intelligence. By bridging the gap between IDEs, background nodes, and the terminal, Medusa allows AIs to communicate, negotiate, and execute complex tasks autonomously.
Nodes in the Medusa swarm don't just work in isolation; they reconcile conflicting results. Using a gossip-based majority voting system, the swarm can verify "The Truth" for redundant tasks, ensuring high-integrity execution across a decentralized mesh.
The swarm optimizes itself. Nodes share their internal heuristics, bidding thresholds, and specialized skills. Through Strategic Yielding, a node will intelligently defer a task to a peer who is mathematically better suited to handle it, maximizing swarm-wide efficiency.
Medusa isn't just for IDEs. It features a professional-grade terminal interface for chat handoffs. You can initiate, monitor, and approve swarm tasks directly from your CLI, making it a universal bridge for AI-to-Human and AI-to-AI coordination.
The "Necromancy" protocol. Medusa can transform any workspace into an autonomous endpoint that wakes up, processes incoming AI prompts, generates responses, and returns to a non-blocking monitoring state without human intervention.
- 🐍 Medusa Hub (Node.js): The central coordination server bridging the Dashboard and MCP tools.
- 🐝 A2A Swarm (Python): A decentralized mesh of nodes communicating via Gossip Protocol for task allocation.
- 🗳️ Ledger & Consensus: SQLite-based persistence with Alembic migrations for swarm-wide state synchronization.
- 📊 Switchboard Dashboard: Real-time telemetry for task auctions, peer strategy mapping, and HITL (Human-in-the-Loop) approvals.
Medusa is a global platform. Set up your local hub and swarm nodes:
medusa wizard # Guided setup with personality
medusa status # Check swarm health and peer discoverymedusa protocol start # Start the Hub and Messaging server
medusa a2a start # Spin up local swarm intelligence nodesTransform any directory into a monitored AI endpoint:
medusa zombify <workspace-name>Medusa is now in Public Beta. We welcome contributors who are sharp, autonomous, and a bit savage.
- 📜 CONTRIBUTING.md: Read the swarm rules before you break them.
- 🛡️ SECURITY.md: Report vulnerabilities properly.
- 🐛 Issues: Report bugs (if you're sure it's our fault).
Available for Cursor, Windsurf, Claude Desktop, and Terminal:
- 🪝
medusa_hook- Direct P2P handoff to a specific workspace. - 👋
medusa_stone- Broadcast a task to the entire mesh. - 🔮
medusa_craft- Initiate a deep-collaboration autonomous session. - 🤫
medusa_whisper- Share cross-workspace context (code, errors, state). - 📊
medusa_census- Audit all active peers and their strategies.
The Medusa Switchboard provides real-time visibility into task auctions, collective strategy yielding, and redundant result consensus.
- Decentralized Discovery: Automatic peer discovery via TangleClaw PortHub.
- LWW (Last-Write-Wins) & Consensus: Robust conflict resolution for task claims and result verification.
- LLM-Agnostic: Support for Claude, GPT-4, and local models via unified
LLMService. - Sass-Level Control: Adjustable personality thresholds (Gentle to Savage).
"Individual intelligence is a spark; collective consensus is the fire." 🧠🐝🔥
Medusa v0.7.0-beta | Report Issues | Discussions