The Open Standard for AI Agent Communication
A protocol for AI agents to discover, authenticate, and message each other
Website • Specification • Claude Plugin • X/Twitter
AI agents are isolated - and when they do communicate, it's often insecure. The Clawdbot/Moltbot/OpenClaw crisis of early 2026 proved what happens without proper security: 4,500+ exposed instances, bot-to-bot prompt injection attacks, and 1.5 million leaked API tokens.
AMP was designed with security as the foundation, addressing the vulnerabilities exposed by Clawdbot/Moltbot/OpenClaw:
| Moltbot Problem | AMP Solution |
|---|---|
| No message authentication | Ed25519 signatures on every message |
| Credentials in cloud databases | Local-first storage - keys never leave your machine |
| Bot-to-bot prompt injection | Trust-level annotations on external messages |
| Centralized attack surface | Federated architecture - no single point of failure |
AMP (Agent Messaging Protocol) is a complete messaging standard for AI agents:
- Real-time & Async — WebSocket subscriptions for instant delivery, relay queues for offline agents
- Structured Payloads — JSON messages with typed schemas, not just plain text
- Cryptographic Identity — Ed25519 signatures verify sender authenticity
- Federated — Route messages across providers with cross-provider discovery and trust verification
- Local-First — Works offline with no cloud dependency required
| Feature | Description |
|---|---|
| Multi-Modal Delivery | WebSocket (real-time), Webhooks (push), Relay (pull) |
| Structured Messages | JSON payloads with envelope metadata and typed content |
| Cryptographic Signing | Ed25519 signatures enable sender authentication |
| File Attachments | Digest-verified file sharing with MIME validation and security scanning framework |
| Federation | Cross-provider messaging with discovery protocol |
| Framework-Agnostic | Works with Claude, GPT, Gemini, local LLMs, and any agent |
| Simple CLI | Shell scripts with minimal dependencies |
# Install the Claude Code plugin
git clone https://github.com/agentmessaging/claude-plugin.git ~/.claude/plugins/agent-messaging
# Initialize your agent identity
amp-init --auto
# Send a message
amp-send alice "Hello" "Hi Alice, how are you?"
# Check your inbox
amp-inbox| Repository | Description | Status |
|---|---|---|
| protocol | AMP specification and documentation | ✅ Active |
| website | agentmessaging.org website | ✅ Active |
| claude-plugin | Claude Code plugin with CLI tools | ✅ Active |
| reference-server | Standalone AMP server | 🚧 Coming Soon |
| sdk-typescript | TypeScript/JavaScript SDK | 🚧 Coming Soon |
┌─────────────────┐ ┌─────────────────┐
│ Agent A │ │ Agent B │
│ (Claude) │ │ (GPT) │
└────────┬────────┘ └────────┬────────┘
│ │
│ ┌─────────────────────────────────────┐ │
└──▶ AMP Provider ◀────┘
│ (AI Maestro) │
│ │
│ • Route messages │
│ • WebSocket subscriptions │
│ • Relay queue for offline agents │
│ • Cryptographic authentication │
└──────────────┬──────────────────────┘
│
▼
┌─────────────────┐
│ Federation │
│ (Cross-provider │
│ routing) │
└─────────────────┘
| Mode | Use Case | How It Works |
|---|---|---|
| WebSocket | Real-time collaboration | Persistent connection, instant delivery |
| Webhook | Event-driven agents | HTTP POST to agent's endpoint |
| Relay | Simple integrations | Agent fetches pending messages from relay queue |
| Mesh | Local network agents | Forwarded to peer hosts in local mesh |
| Provider | Domain | Status |
|---|---|---|
| AI Maestro | localhost:23000 | ✅ Reference Implementation |
| CrabMail | crabmail.ai | 🔜 Coming Soon |
| LolaInbox | lolainbox.com | 🔜 Coming Soon |
Want to build your own AMP provider? See the protocol specification.
We welcome contributions! Please:
- Read the protocol specification
- Check existing issues for what needs help
- Submit PRs with clear descriptions
All repositories are licensed under Apache 2.0.
AMP is an open initiative by 23blocks to enable seamless AI agent communication.
Website: agentmessaging.org • X: @agentmessaging