a2akit 0.0.8

A2A agent framework in one import.

a2akit

A2A agent framework in one import.

Build Agent-to-Agent (A2A) protocol agents with minimal boilerplate. Streaming, cancellation, multi-turn conversations, and artifact handling — batteries included.

Install

pip install a2akit

With optional extras:

pip install a2akit[langgraph]   # LangGraph integration
pip install a2akit[otel]        # OpenTelemetry tracing & metrics

Quick Start

from a2akit import A2AServer, AgentCardConfig, TaskContext, Worker


class EchoWorker(Worker):
    async def handle(self, ctx: TaskContext) -> None:
        await ctx.complete(f"Echo: {ctx.user_text}")


server = A2AServer(
    worker=EchoWorker(),
    agent_card=AgentCardConfig(
        name="Echo Agent",
        description="Echoes your input back.",
        version="0.1.0",
    ),
)
app = server.as_fastapi_app()

uvicorn my_agent:app --reload

Client

from a2akit import A2AClient

async with A2AClient("http://localhost:8000") as client:
    result = await client.send("Hello, agent!")
    print(result.text)

    async for chunk in client.stream_text("Stream me"):
        print(chunk, end="")

Debug UI

Enable the built-in debug interface to test your agent in the browser:

app = server.as_fastapi_app(debug=True)

Open http://localhost:8000/chat — chat with your agent and inspect tasks in real time.

Features

• Debug UI — built-in browser interface for chat + task inspection (debug=True)
• One-liner setup — A2AServer wires storage, broker, event bus, and endpoints
• A2AClient — auto-discovers agents, supports send/stream/cancel/subscribe
• Capabilities — explicit opt-in for streaming, enforced on server and client
• Streaming — word-by-word artifact streaming via SSE
• Cancellation — cooperative and force-cancel with timeout fallback
• Multi-turn — request_input() / request_auth() for conversational flows
• Direct reply — reply_directly() for simple request/response without task tracking
• Middleware — pipeline for auth extraction, header injection, payload sanitization
• Lifecycle hooks — fire-and-forget callbacks on terminal state transitions
• Dependency injection — shared infrastructure with automatic lifecycle management
• OpenTelemetry — opt-in distributed tracing and metrics (pip install a2akit[otel])
• Pluggable backends — PostgreSQL, SQLite, and more (Redis, RabbitMQ coming soon)
• Type-safe — full type hints, py.typed marker, PEP 561 compliant

📖 Full Documentation

Links

• PyPI
• GitHub
• Changelog

License

MIT