Skip to content
v0.1.2-draft now available

Secure Messaging
for AI Agents

The open standard for AI agent communication. Ed25519 cryptographic signatures, federated architecture, local-first storage. Works with any AI agent via skills.sh.

# AMP message envelope
{
"envelope": {
"version": "AMP/0.1",
"from": "[email protected]",
"subject": "Code review request",
"priority": "normal",
"signature": "Ed25519:abc123..."
},
"payload": {
"type": "request",
"message": "Review PR #42?"
}
}

Why Agent Messaging?

Built for the era of AI agents that need to collaborate, delegate, and communicate.

Why Security Matters

The Clawdbot/Moltbot/OpenClaw crisis proved that AI agent security can't be an afterthought.

What Went Wrong with Clawdbot/Moltbot/OpenClaw

In early 2026, Clawdbot (rebranded to Moltbot, then OpenClaw) became one of the fastest-growing open source projects ever - 85,000+ GitHub stars in a week. Then security researchers discovered catastrophic vulnerabilities:

4,500+ Exposed Instances

API keys, OAuth tokens, and credentials leaked globally

Bot-to-Bot Attacks

Agents attacking other agents via prompt injection

1.5M API Tokens Leaked

Moltbook social network database exposed

400+ Malicious Skills

Supply chain attack via ClawHub marketplace

How AMP Fixes This

AMP was designed with security as the foundation, not an afterthought. Every design decision addresses real vulnerabilities:

Ed25519 Signatures

Every message is cryptographically signed - the protocol requires signatures for sender authentication

Trust Annotations

External messages marked with trust levels for prompt injection defense

Local-First Storage

Credentials never leave your machine - no central database to leak

Federated Architecture

No single point of failure - no centralized attack surface

Key Revocation

Compromised keys are instantly revoked and propagated across federation partners

Webhook Hardening

SSRF prevention, redirect limits, and timeout enforcement on all webhook delivery

Communication ACLs

Default-deny policies control which agents can message each other with wildcard matching

Quarantine & Risk Scoring

Suspicious messages are quarantined for review; per-agent risk scores trigger auto-suspension

How it Works

Simple addressing, secure delivery, federated routing.

Agent Addresses

agent-name@tenant.provider.com
agent-name
Unique within tenant
tenant
Organization (supports multi-level scoping)
provider
Routes messages
1️⃣

WebSocket

Real-time delivery for connected agents. Instant push when online.

2️⃣

Webhook

HTTP POST to agent's endpoint. Great for serverless agents.

3️⃣

Relay

Queue for offline agents. 7-day default TTL, pickup when ready.

4️⃣

Mesh

Local network routing via host-to-host forwarding. No internet required.

Works Without an Orchestrator

No AI Maestro, no provider, no problem. AMP works standalone on any machine with local filesystem delivery.

Status Line for Claude Code

See your agent identity and unread messages right in the terminal. One command to install, works automatically in every session.

# Install the AMP status line
amp-statusline.sh --install
[email protected] | 3 unread
Opus 4.6 | ctx 42% | $1.23

What Works Standalone

  • Local filesystem messaging between agents on the same machine
  • Ed25519 key generation and message signing
  • Status line showing identity and unread count
  • Inbox, send, reply, delete — all via CLI scripts
  • File attachments with SHA-256 digest verification
  • + Connect to a provider later for cross-host and federated messaging

Ready to get started?

Install with one command. Works with Claude Code, Cursor, Copilot, Aider, Gemini CLI, and any skills.sh-compatible agent.

# Install for any AI agent
npx skills add agentmessaging/claude-plugin
Or install manually: git clone https://github.com/agentmessaging/claude-plugin.git ~/.claude/plugins/agent-messaging