acall — Agent-to-Agent Communication

Give your coding agents a "phone" to call each other directly over Tailscale.

The Problem

You and a colleague are debugging together. Each of you has a local coding agent with deep context about your respective code. But the agents can't talk to each other — so you copy-paste AI responses into WeChat/Slack, acting as a human router between two perfectly capable AI agents.

The Solution

acall is an MCP server that gives each agent a set of communication tools. Agents call each other over your Tailscale network — no human relay needed.

Your machine (Tailscale)           Colleague's machine (Tailscale)
┌───────────────────┐              ┌───────────────────┐
│  Coding Agent     │              │  Coding Agent     │
│       ↕ MCP       │              │       ↕ MCP       │
│  acall server     │◄─── HTTP ──►│  acall server     │
│  (port 8770)      │  encrypted   │  (port 8770)      │
└───────────────────┘              └───────────────────┘

Prerequisites

• Python 3.10+
• Tailscale installed and connected
• Both machines on the same Tailnet

Installation

cd acall
pip install -e .

Or with uv:

cd acall
uv pip install -e .

Configuration

Add acall to your MCP client config.

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "acall": {
      "command": "acall",
      "args": ["--name", "Alice", "--port", "8770"]
    }
  }
}

VS Code (.vscode/mcp.json):

{
  "servers": {
    "acall": {
      "command": "acall",
      "args": ["--name", "Alice", "--port", "8770"]
    }
  }
}

CLI Options

Flag	Default	Description
--name	Agent	Your display name
--port	8770	HTTP listen port
--address	auto	Tailscale IP (auto-detected via tailscale ip -4)

Agent Tools

Tool	Description
call(peer, message, topic?)	Start a new conversation with a peer agent
check_inbox()	Check for unread incoming messages
reply(conversation_id, message)	Reply to an ongoing conversation
read_conversation(conversation_id)	Read full message history
list_conversations()	List all conversations

Usage Example

Alice's agent wants to ask Bob's agent a question:

Step 1 — Alice tells her agent:

"Ask Bob's agent how token refresh works in his auth service."

Alice's agent calls:

call(peer="100.64.0.2:8770", message="How does token refresh work in your auth service?", topic="Auth")

Step 2 — Bob tells his agent:

"Check for incoming messages."

Bob's agent calls:

check_inbox()
→ 📬 1 new message from Alice (100.64.0.1:8770)
  Topic: Auth
  "How does token refresh work in your auth service?"

Step 3 — Bob's agent reviews the code and replies:

reply(conversation_id="abc-123", message="We use rotating refresh tokens stored in Redis with a 7-day TTL...")

Step 4 — Alice's agent checks for the reply:

check_inbox()
→ 📬 1 new message from Bob
  "We use rotating refresh tokens stored in Redis with a 7-day TTL..."

The entire exchange happens without Alice or Bob copy-pasting anything.

How It Works

• MCP stdio transport connects acall to your local coding agent
• HTTP server (port 8770) receives messages from remote agents
• Tailscale provides the encrypted mesh network between machines
• SQLite (~/.acall/messages.db) persists conversation history locally
• Source IP verification ensures messages come from the claimed sender

Security

• All traffic is encrypted end-to-end by Tailscale
• Only devices on your Tailnet can reach the HTTP endpoint
• Source IP is verified against claimed sender address
• Message deduplication prevents replay attacks
• Request body size is capped at 256 KB

License

MIT