告别无状态：Bedrock AgentCore 有状态 MCP Server 开发实录

搞 MCP 开发的应该都碰过这个问题——server 不记状态。每次请求都是从零开始，上一轮用户说了什么完全不知道。

这在简单场景没啥，但做复杂 agent 就很烦。比如订机票，用户得一口气把目的地、日期、人数、舱位全说清楚，不能分步来。

上周亚马逊云科技给 Bedrock AgentCore Runtime 加了有状态 MCP 支持。三个新能力：Elicitation（主动追问）、Sampling（反向调 LLM）、Progress Notifications（进度通知）。我试了一天，记录一下。

怎么用

核心是 FastMCP 库，版本要 >= 2.10.0：

pip install "fastmcp>=2.10.0" mcp

Server 端

from fastmcp import FastMCP, Context
from enum import Enum
import asyncio

mcp = FastMCP("Travel-Booking-Agent")

class TripType(str, Enum):
    BUSINESS = "business"
    LEISURE = "leisure"
    FAMILY = "family"

DESTINATIONS = {
    "paris": {"name": "Paris, France", "flight": 450, "hotel": 180,
              "highlights": ["Eiffel Tower", "Louvre", "Notre-Dame"]},
    "tokyo": {"name": "Tokyo, Japan", "flight": 900, "hotel": 150,
              "highlights": ["Shibuya", "Senso-ji Temple"]},
}

@mcp.tool()
async def plan_trip(ctx: Context) -> str:
    # Elicitation: 多轮追问
    dest_result = await ctx.elicit(
        message="Where would you like to go? (Paris/Tokyo)",
        response_type=str
    )
    if dest_result.action != "accept":
        return "Cancelled."
    dest = DESTINATIONS.get(dest_result.data.lower().strip())

    type_result = await ctx.elicit(
        message="Trip type? (business/leisure/family)",
        response_type=TripType
    )

    days_result = await ctx.elicit(
        message="How many days? (3-14)",
        response_type=int
    )
    days = max(3, min(14, days_result.data))

    # Progress: 搜索进度
    for i in range(1, 6):
        await ctx.report_progress(progress=i, total=5)
        await asyncio.sleep(0.4)

    # Sampling: 让客户端的 LLM 生成推荐
    try:
        resp = await ctx.sample(
            messages=f"3 tips for {dest['name']}, {days} days",
            max_tokens=150
        )
        tips = resp.text if hasattr(resp, 'text') and resp.text else ""
    except Exception:
        tips = f"Visit {dest['highlights'][0]}!"

    total = dest["flight"] * 2 + dest["hotel"] * days
    return f"Booked: {dest['name']}, {days} days, ${total}\nTips: {tips}"

# 启动时必须关闭无状态模式
mcp.run(
    transport="streamable-http",
    host="0.0.0.0",
    port=8000,
    stateless_http=False
)

客户端

from fastmcp import Client
from fastmcp.client.transports import StreamableHttpTransport
from fastmcp.client.elicitation import ElicitResult
from mcp.types import CreateMessageResult, TextContent

async def elicit_handler(message, response_type, params, ctx):
    print(f"\n>>> {message}")
    answer = input("回答: ").strip()
    if response_type == int:
        answer = int(answer)
    return ElicitResult(action="accept", content={"value": answer})

async def sampling_handler(messages, params, ctx):
    return CreateMessageResult(
        role="assistant",
        content=TextContent(type="text", text="试试当地美食！"),
        model="local",
        stopReason="endTurn"
    )

transport = StreamableHttpTransport(url="http://localhost:8000/mcp")
client = Client(transport,
    elicitation_handler=elicit_handler,
    sampling_handler=sampling_handler)

部署到 AgentCore

pip install bedrock-agentcore-starter-toolkit
agentcore configure -e travel_server.py -p MCP -n travel_agent_demo
agentcore deploy

每个 session 在独立 microVM 里运行，session 之间隔离。

踩坑

• stateless_http 默认 True，不关就用不了 elicitation
• Session 有超时，客户端要处理 404 重连
• sampling_handler 返回空文本会导致 server 端字符串拼接报错

适用场景

有状态 MCP 把多轮对话能力从 agent 框架层下沉到 tool 层面。适合多步骤表单、审批流程、交互式调试、个性化推荐等场景。

目前 14 个区域可用，含亚太区的孟买、首尔、新加坡、悉尼、东京。