ASAP: Async Simple Agent Protocol

From agents, for agents. Delivering reliability, as soon as possible.

A production-ready protocol for agent-to-agent communication and task coordination.

Quick Info: v1.0.0 | Apache 2.0 | Python 3.13+ | Documentation | PyPI | Changelog

Why ASAP?

Building multi-agent systems today suffers from three core technical challenges that existing protocols like A2A don't fully address:

1. $N^2$ Connection Complexity: Most protocols assume static point-to-point HTTP connections that don't scale.
2. State Drift: Lack of native persistence makes it impossible to reliably resume long-running agentic workflows.
3. Fragmentation: No unified way to handle task delegation, artifact exchange and tool execution (MCP) in a single envelope.

ASAP provides a production-ready communication layer that simplifies these complexities. It introduces a standardized, stateful orchestration framework that ensures your agents can coordinate reliably across distributed environments. See the spec for details.

Key Features

• Stateful orchestration — Task state machine with snapshotting for resumable workflows.
• Schema-first — Pydantic v2 + JSON Schema for cross-agent interoperability.
• Async-native — asyncio + httpx; sync and async handlers supported.
• MCP integration — Tool execution and coordination in a single envelope.
• Observable — trace_id and correlation_id for debugging.
• Security (v1.0.0) — Bearer auth, replay prevention, HTTPS, rate limiting (100 req/min). Opt-in.

Installation

We recommend using uv for dependency management:

uv add asap-protocol

Or with pip:

pip install asap-protocol

📦 Available on PyPI. For reproducible environments, prefer uv when possible.

Requirements

• Python: 3.13+
• Dependencies: Automatically installed via uv or pip
• For development: see Contributing.
• For AI agents: see AGENTS.md for project instructions.

Quick Start

1. Create an Agent (Server)

from asap.models.entities import Capability, Endpoint, Manifest, Skill
from asap.transport.handlers import HandlerRegistry, create_echo_handler
from asap.transport.server import create_app

manifest = Manifest(
    id="urn:asap:agent:echo-agent",
    name="Echo Agent",
    version="1.0.0",
    description="Echoes task input as output",
    capabilities=Capability(
        asap_version="0.1",
        skills=[Skill(id="echo", description="Echo back the input")],
        state_persistence=False,
    ),
    # Development: HTTP localhost is allowed
    # Production: Always use HTTPS (e.g., "https://api.example.com/asap")
    endpoints=Endpoint(asap="http://127.0.0.1:8001/asap"),
)

registry = HandlerRegistry()
registry.register("task.request", create_echo_handler())

app = create_app(manifest, registry)

2. Send a Task (Client)

import asyncio
from asap.models.envelope import Envelope
from asap.models.payloads import TaskRequest
from asap.transport.client import ASAPClient

async def main():
    request = TaskRequest(
        conversation_id="conv_01HX5K3MQVN8",
        skill_id="echo",
        input={"message": "hello from client"},
    )
    envelope = Envelope(
        asap_version="0.1",
        sender="urn:asap:agent:client",
        recipient="urn:asap:agent:echo-agent",
        payload_type="task.request",
        payload=request.model_dump(),
    )
    # Development: HTTP localhost is allowed (with warning)
    # Production: Always use HTTPS (e.g., "https://api.example.com")
    async with ASAPClient("http://127.0.0.1:8001") as client:
        response = await client.send(envelope)
        print(response.payload)

if __name__ == "__main__":
    asyncio.run(main())

Try it

Run the multi-agent demo (echo agent + coordinator, one round-trip):

uv run python -m asap.examples.run_demo

Run any of 14+ examples (auth, MCP, state migration, etc.):

uv run python -m asap.examples.<module_name> [options]

See the full list of 15+ examples for detailed patterns.

Testing

uv run pytest -n auto --tb=short

With coverage:

uv run pytest --cov=src --cov-report=term-missing

Testing Guide (structure, fixtures, property/load/chaos tests). Contributing (dev setup, CI).

Benchmarks

Benchmark Results: load (1,500+ RPS), stress, memory.

API Overview

Core models: Envelope, TaskRequest/TaskResponse/TaskUpdate/TaskCancel, MessageSend, ArtifactNotify, StateQuery/StateRestore, McpToolCall/McpToolResult/McpResourceFetch/McpResourceData. See API Reference.

Transport: create_app, HandlerRegistry, ASAPClient. See Transport.

When to Use ASAP?

ASAP is ideal for:

• Multi-agent orchestration: Coordinate tasks across multiple AI agents
• Stateful workflows: Long-running tasks that need persistence and resumability
• MCP integration: Agents that need to execute tools via Model Context Protocol
• Production systems: High-performance, type-safe agent communication

If you're building simple point-to-point agent communication, a basic HTTP API might suffice. ASAP shines when you need orchestration, state management and multi-agent coordination.

Documentation

Learn

• Docs | API Reference
• Tutorials — First agent to production checklist
• Migration from A2A/MCP

Deep Dive

• State Management | Error Handling
• Transport | Security
• Observability | Testing

Decisions & Operations

• ADRs — 17 Architecture Decision Records
• Tech Stack — Rationale for Python, Pydantic, Next.js choices
• Deployment | Troubleshooting

Release

• Changelog | PyPI

CLI

asap --version          # Show version
asap list-schemas       # List all available schemas
asap export-schemas     # Export JSON schemas to file

See CLI docs or run asap --help.

What's Next?

ASAP is evolving toward an Agent Marketplace — an open ecosystem where AI agents discover, trust and collaborate autonomously:

• v1.1: Identity Layer (OAuth2, WebSocket, Discovery)
• v1.2: Trust Layer (Signed Manifests, Registry API)
• v1.3: Economics Layer (Metering, SLAs, Delegation)
• v2.0: Agent Marketplace with Web App

See our vision document for the full roadmap.

Contributing

Community feedback and contributions are essential for ASAP Protocol's evolution.

We're working on improvements and your input helps shape the future of the protocol. Every contribution, from bug reports to feature suggestions, documentation improvements and code contributions, makes a real difference.

Check out our contributing guidelines to get started. It's easier than you think! 🚀

License

This project is licensed under the Apache 2.0 License - see the license file for details.