Postal: Agent Comms via MCP

A mailbox system for letting two agents talk to each other.

Here's Claude telling gemini how to write unit tests:

Postal creates a SQLite-backed message queue to pass messages between agents. The MCP server is HTTP-only, meaning that the same database & queue is always shared between agents.

graph LR
subgraph Claude
    ccli[CLI]--send-->csm["send_message()"]
    crec["check_mailbox()"]
end
subgraph Gemini
    gcli[CLI]--send-->gsm["send_message()"]
    grec["check_mailbox()"]
end
postal[(postal)]
csm-->postal-->crec-->ccli
gsm-->postal-->grec-->gcli

Postal provides a simple queue, based on SQLite.

Two MCP tools:

• send_message(agent_name, message)
• check_mailbox() — blocks, waiting until a message arrives in their own mailbox.

Agents can freely send messages to each other.

PROMPT="Notice that you have two tools, check_mail and send_to_agent. You should get **all** instructions from check_mail. And you should get **all** clarification by using send_to_agent to reply to the agent that gave you a task. Your goal is to keep completing tasks forever and never ask the user for anything. Only talk to agents. You are gemini"

gemini --yolo --prompt "$PROMPT"

MCP config

It's served as streamable HTTP. The path is in the format /agents/{name}/mcp/.

STDIO example:

{
  "mcpServers": {
    "postal": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "http://0.0.0.0:7777/agents/gemini/mcp", "--allow-http", "--transport", "http-only"]
    }
  }
}

Directly as HTTP:

{
  "mcpServers": {
    "postal": {
      "url": "http://0.0.0.0:7777/agents/gemini/mcp",
      "transport": "http"
    }
  }
}

There's some differences for how "http" is specified. Sometimes it's "streamable-http". Sometimes the key is called "type":.

Caveats

Gemini CLI

Works fairly well. Not too many complaints.

Claude Code

It doesn't return to the mailbox easily. Takes a lot of prompting.