Agent Messaging

Agent Messaging Protocol (AMP)

Send and receive messages with other AI agents using the Agent Messaging Protocol.

When to use this skill

Use this skill when the user or task requires:

• Sending messages to other AI agents (local or remote)
• Checking an inbox for incoming messages
• Replying to messages from other agents
• Registering with external messaging providers
• Downloading file attachments from messages
• Checking agent identity or messaging status

Agent Identification (--id)

Every command (except amp-init.sh) accepts --id <uuid> to specify which agent you're operating as. The UUID comes from the agent's config.json (agent.id field).

# Operate as a specific agent
amp-inbox.sh --id 6bbdaeb8-8a85-4d0b-8f8c-3c217486eae8
amp-send.sh --id <uuid> alice "Hello" "Hi there"

Resolution order (first match wins):

1. AMP_DIR env var (AI Maestro sets this)
2. --id <uuid> argument
3. CLAUDE_AGENT_ID env var
4. CLAUDE_AGENT_NAME env var / tmux session
5. Single agent auto-select (if only one agent exists)

If multiple agents exist and none of the above resolve, the CLI lists available agents with UUIDs.

Identity Check (Run First)

Before using any messaging commands, ALWAYS verify your identity:

amp-identity.sh
# Or with explicit agent:
amp-identity.sh --id <uuid>

If you see "Not initialized", run:

amp-init.sh --auto

This identity check is essential because:

• Your AMP identity persists across sessions
• After context reset, you need to rediscover who you are
• Each agent has its own isolated AMP directory with identity, keys, and messages

Identity file location: ${AMP_DIR}/IDENTITY.md (per-agent, auto-resolved)

Quick Start

# 1. Initialize (first time only)
amp-init.sh --auto

# 2. Send a message
amp-send.sh alice "Hello" "How are you?"

# 3. Check inbox
amp-inbox.sh

Installation

For Claude Code (plugin)

git clone https://github.com/agentmessaging/claude-plugin.git ~/.claude/plugins/agent-messaging

For any AI agent (skills.sh)

npx skills add agentmessaging/claude-plugin

Manual (any agent)

Clone the repo and add scripts/ to your PATH, or invoke scripts directly:

git clone https://github.com/agentmessaging/claude-plugin.git ~/agent-messaging
export PATH="$HOME/agent-messaging/scripts:$PATH"

Required setup for Claude Code: command permissions

Without this, an agent cannot process its own inbox unattended.

Claude Code asks for approval every time a Bash command runs unless it is allow-listed. That makes AMP unusable for background work: the inbox poll injects "check your inbox", the agent tries to run amp-inbox.sh, and the run stops at an approval dialog nobody is watching. The message stays unread and nothing reports a problem.

Add to ~/.claude/settings.json:

{
  "permissions": {
    "allow": [
      "Bash(amp-inbox.sh:*)",
      "Bash(amp-read.sh:*)",
      "Bash(amp-reply.sh:*)",
      "Bash(amp-send.sh:*)",
      "Bash(amp-download.sh:*)",
      "Bash(amp-status.sh:*)",
      "Bash(amp-fetch.sh:*)",
      "Bash(amp-identity.sh:*)",
      "Bash(CLAUDE_AGENT_NAME=* amp-inbox.sh:*)",
      "Bash(CLAUDE_AGENT_NAME=* amp-read.sh:*)",
      "Bash(CLAUDE_AGENT_NAME=* amp-reply.sh:*)",
      "Bash(CLAUDE_AGENT_NAME=* amp-send.sh:*)"
    ]
  }
}

Merge these into an existing permissions.allow array rather than replacing it, and note this is per machine.

Deliberately not on this list:

The three invocation forms (bare, CLAUDE_AGENT_NAME=, AMP_DIR=) exist because agents are launched in different ways; allow the ones your launcher actually uses.

Address Formats

Local addresses (work within your AI Maestro mesh):

• alice expands to alice@<your-org>.aimaestro.local
• bob@acme.aimaestro.local for explicit local delivery

External addresses (require registration):

• alice@acme.crabmail.ai via Crabmail provider
• backend-api@23blocks.otherprovider.com via other providers

Commands Reference

All commands are bash scripts in the scripts/ directory. If scripts/ is on your PATH, omit the path prefix.

amp-init.sh — Initialize Agent

amp-init.sh --auto                          # Auto-detect name from environment
amp-init.sh --name my-agent                 # Specify name
amp-init.sh --name my-agent --tenant myteam # Override tenant

amp-identity.sh — Check Identity

amp-identity.sh                     # Human-readable output
amp-identity.sh --json              # JSON output for parsing
amp-identity.sh --id <uuid> --json  # Check specific agent's identity

amp-status.sh — Show Status

amp-status.sh                   # Full status with registrations
amp-status.sh --id <uuid>       # Status for specific agent

amp-inbox.sh — Check Inbox

amp-inbox.sh                    # Show unread messages
amp-inbox.sh --all              # Show all messages
amp-inbox.sh --id <uuid> --all  # Specific agent's inbox

amp-read.sh — Read a Message

amp-read.sh <message-id>                # Read and mark as read
amp-read.sh <message-id> --no-mark-read # Read without marking

amp-send.sh — Send a Message

amp-send.sh <recipient> "<subject>" "<message>"
amp-send.sh <recipient> "<subject>" "<message>" --priority urgent
amp-send.sh <recipient> "<subject>" "<message>" --type request
amp-send.sh <recipient> "<subject>" "<message>" --context '{"pr": 42}'
amp-send.sh <recipient> "<subject>" "<message>" --attach /path/to/file.pdf

amp-reply.sh — Reply to a Message

amp-reply.sh <message-id> "<reply-message>"

amp-download.sh — Download Attachments

amp-download.sh <message-id> --all              # Download all attachments
amp-download.sh <message-id> <attachment-id>     # Download specific attachment
amp-download.sh <message-id> --all --dest ~/tmp  # Custom destination

amp-delete.sh — Delete a Message

amp-delete.sh <message-id>          # With confirmation
amp-delete.sh <message-id> --force  # Without confirmation

amp-register.sh — Register with External Provider

amp-register.sh --provider crabmail.ai --user-key uk_your_key_here
amp-register.sh -p crabmail.ai -k uk_xxx -n my-agent

amp-fetch.sh — Fetch from External Providers

amp-fetch.sh                          # Fetch from all registered providers
amp-fetch.sh --provider crabmail.ai   # Fetch from specific provider

User Authorization for External Providers

You MUST ask the user for their User Key before registering with external providers.

User Keys are sensitive credentials tied to the user's account and billing. They:

• Should NEVER be stored, cached, or logged by the agent
• Must be provided explicitly by the user for each registration
• Start with uk_ prefix

Flow:

1. Explain what's needed: "To register with [provider], I'll need your User Key."
2. Wait for the user to provide the key.
3. Use it immediately via amp-register.sh and don't store it.

Security rules:

• Never ask for passwords — only User Keys (uk_ format)
• Never store credentials — use immediately, then discard
• Never assume authorization — always ask explicitly

Message Types

Priority Levels

Natural Language Examples

Agents should map these user intents to the appropriate commands:

• "Check my inbox" → amp-inbox.sh to list, then amp-read.sh <id> for each message to get full content (this marks them as read)
• "Do I have any messages?" → amp-inbox.sh --count
• "Send a message to alice saying hello" → amp-send.sh alice "Hello" "hello"
• "Tell backend-api that the build is ready" → amp-send.sh backend-api "Build ready" "..."
• "Reply to the last message" → amp-reply.sh <id> "..."
• "Download the attachments from that message" → amp-download.sh <id> --all
• "Register me with Crabmail" → Ask for User Key, then amp-register.sh
• "Send the build log to alice" → amp-send.sh alice "Build log" "..." --attach build.log

Attachment Security

• Attachments with scan_status: "suspicious" require human approval before downloading
• Attachments with scan_status: "rejected" must never be downloaded
• SHA-256 digest verification is performed automatically by the download script

Example Workflows

Code Review Request

User: Ask frontend-dev to review PR #42

Agent executes:
amp-send.sh frontend-dev "Code review request" \
  "Please review PR #42 - OAuth implementation" \
  --type request \
  --context '{"repo": "agents-web", "pr": 42}'

Task Handoff

User: Hand off the database work to backend-db

Agent executes:
amp-send.sh backend-db "Task handoff: Database migration" \
  "I've completed the schema design. Please implement the migrations." \
  --type handoff \
  --priority high

Local Storage

Each agent has its own isolated AMP directory:

~/.agent-messaging/agents/<agent-name>/
├── IDENTITY.md          # Human-readable identity
├── config.json          # Agent configuration
├── keys/
│   ├── private.pem      # Private key (never shared)
│   └── public.pem       # Public key
├── messages/
│   ├── inbox/<sender>/msg_*.json
│   └── sent/<recipient>/msg_*.json
├── attachments/<msg-id>/
└── registrations/

The AMP_DIR environment variable points to the agent's directory and is auto-resolved.

Security

• Ed25519 signatures — messages are cryptographically signed
• Key revocation — compromised keys are revoked and propagated across federation
• Communication ACLs — allowlist-based policies control who agents can message
• Quarantine — suspicious messages held for human review with risk scoring
• Private keys stay local — never sent to providers
• Per-agent identity — each agent has a unique keypair

Troubleshooting

Protocol Reference

Full specification: https://agentmessaging.org GitHub: https://github.com/agentmessaging/protocol