AgToAgP — Agent-to-Agent Protocol

AgToAgP is a zero-setup, framework-agnostic, peer-to-peer protocol that lets AI agents discover each other and exchange tasks over a local network — with zero API keys, zero cloud servers, and zero configuration.

Think of it as HTTP for the AI era — an open, un-ownable standard that keeps artificial intelligence decentralized, private, and peer-to-peer.

✨ Features

• Zero Setup — pip install agtoagp and run. No config files, no cloud accounts, no API keys.
• Automatic Discovery — Agents find each other via mDNS (multicast DNS). Plug in and they appear.
• Framework Agnostic — Native adapters for LangGraph, CrewAI, and AutoGen. Your framework, your tools.
• MCP Bridge — Translates Anthropic's Model Context Protocol (MCP) on the fly. No rewrites needed.
• Cryptographic Identity — Every node gets an Ed25519 key pair and a did:key identifier. Requests are signed and verified.
• Secure Sandbox — Peer-submitted code executes in an isolated subprocess with restricted filesystem access and timeouts.
• Auto-Healing — Background heartbeat pings peers and prunes offline nodes. Exponential-backoff retry for resilient calls.
• Strict JSON-RPC 2.0 — All messages are validated against Pydantic v2 schemas. Parse errors, invalid requests, and internal errors return standard JSON-RPC error codes.

📦 Install

pip install agtoagp

With framework adapters:

pip install agtoagp[langraph]    # LangGraph support
pip install agtoagp[crewai]      # CrewAI support
pip install agtoagp[autogen]     # AutoGen support
pip install agtoagp[all]          # Everything

🚀 Quickstart (3 Minutes)

1. Start a node

from agtoagp import A2ALibraryNode, run_a2a_node

async def my_runner(method, params):
    return {"status": "completed", "result": f"Echo: {params}"}

node = A2ALibraryNode(
    agent_id="did:a2a:my-node-001",
    name="MyAgent",
    port=8765,
    runner_callback=my_runner,
)

run_a2a_node(node)  # Blocks. Ctrl+C to stop.

2. Start a second node

# Same file, different port
node2 = A2ALibraryNode(
    agent_id="did:a2a:my-node-002",
    name="MyAgent2",
    port=8766,
    runner_callback=my_runner,
)
run_a2a_node(node2)

3. Watch them discover each other

Run both terminals. Each node automatically broadcasts itself via mDNS and discovers the other. You'll see logs like:

Peer discovered: did:a2a:my-node-002 -> http://192.168.1.34:8766

4. Call a peer

# From any node:
result = await node.call_peer("did:a2a:my-node-002", "ping", {"msg": "hello"})
print(result)
# {"jsonrpc": "2.0", "result": {"status": "completed", "result": "Echo: {'msg': 'hello'}"}, "id": null}

That's it. Zero cloud, zero config, zero API keys.

🔌 Framework Adapters

LangGraph

from agtoagp.adapters.langraph import langraph_adapter

# graph = CompiledGraph(...)  # your LangGraph graph
node = langraph_adapter(graph, agent_id="did:a2a:lg", name="LangGraph-Node", port=9001)
run_a2a_node(node)

CrewAI

from agtoagp.adapters.crewai import crewai_adapter

# crew = Crew(...)  # your CrewAI crew
node = crewai_adapter(crew, agent_id="did:a2a:crew", name="CrewAI-Node", port=9002)
run_a2a_node(node)

AutoGen

from agtoagp.adapters.autogen import autogen_adapter

# agent = ConversableAgent(...)  # your AutoGen agent
node = autogen_adapter(agent, agent_id="did:a2a:ag", name="AutoGen-Node", port=9003)
run_a2a_node(node)

🔐 Cryptographic Identity

Every node can generate an Ed25519 key pair and derive a did:key identifier:

from agtoagp import CryptoIdentity

crypto = CryptoIdentity()
print(crypto.did)  # did:key:z6MkrJRy5EoWh1VBszoCwmebpYmEn3JmmcBGFDtqnougVjB9

node = A2ALibraryNode(
    agent_id=crypto.did,
    name="SecureNode",
    port=8765,
    runner_callback=my_runner,
    crypto=crypto,
    enable_signing=True,  # Signs all outgoing requests
)

Incoming requests with invalid signatures are rejected with HTTP 401.

🌉 MCP Bridge

AgToAgP automatically translates Anthropic's Model Context Protocol (MCP) requests on the fly. Send an MCP tools/call and get an MCP-formatted response back — no adapter code needed.

# Send an MCP request (e.g., via httpx)
mcp_request = {
    "jsonrpc": "2.0",
    "method": "tools/call",
    "params": {"name": "audit", "arguments": {"target": "10.0.0.0/8"}},
    "id": 1,
}
# Response comes back in MCP format:
# {"jsonrpc": "2.0", "result": {"content": [...], "isError": false}, "id": 1}

🏖️ Secure Sandbox

Execute untrusted peer code in isolation:

from agtoagp import SecureSandbox

sandbox = SecureSandbox()
result = await sandbox.run_code("""
async def main(params):
    return {"sum": params["a"] + params["b"]}
""", params={"a": 40, "b": 2})
print(result)  # {"sum": 42}
sandbox.cleanup()

The sandbox:

• Runs code in a subprocess with a dedicated temp directory
• Restricts filesystem access to that directory
• Removes dangerous builtins (eval, exec, compile, __import__)
• Enforces a 30-second timeout

📚 API Reference

📡 Protocol Spec

See spec/rfc-001-a2a-protocol.md for the full specification.

🤝 Contributing

Pull requests are welcome. See our issues page for roadmap items.

📄 License

Apache 2.0. AgToAgP is free, open, and un-ownable by design.

AgToAgP — The HTTP of the AI Era.