Build a WeChat AI bot users actually trust

Works with every WeChat surface

What makes WeChat bots hard (and how we fix it)

公众号 Official Account example

User sends text to your 公众号 → WeChat POSTs to your server → you call AIPower → reply via 客服消息 API within 5 seconds.

// Node.js — WeChat Official Account webhook
import OpenAI from "openai";
import xml2js from "xml2js";

const aipower = new OpenAI({
  baseURL: "https://api.aipower.me/v1",
  apiKey: process.env.AIPOWER_API_KEY,
});

export async function handleWeChatMessage(req, res) {
  // Parse WeChat XML POST
  const body = await readBody(req);
  const parsed = await xml2js.parseStringPromise(body);
  const msg = parsed.xml;

  const userId = msg.FromUserName[0];
  const text = msg.Content?.[0];
  if (!text) return res.status(200).send("");

  // Quick reply with AIPower — use auto-fast for <2s latency
  const completion = await aipower.chat.completions.create({
    model: "qwen/qwen-turbo",  // fastest for Chinese text, <500ms first token
    messages: [
      { role: "system", content: "你是公众号客服助手。回答简短（2-3句），必要时引导用户去菜单或官网。" },
      { role: "user", content: text },
    ],
    max_tokens: 200,
    user: userId,
  });

  const reply = completion.choices[0].message.content;

  // Build WeChat XML response
  const xmlReply = `<xml>
<ToUserName><![CDATA[${userId}]]></ToUserName>
<FromUserName><![CDATA[${msg.ToUserName[0]}]]></FromUserName>
<CreateTime>${Math.floor(Date.now()/1000)}</CreateTime>
<MsgType><![CDATA[text]]></MsgType>
<Content><![CDATA[${reply}]]></Content>
</xml>`;
  res.setHeader("Content-Type", "application/xml");
  res.status(200).send(xmlReply);
}

⚠️ WeChat's 5-second response deadline — always use fast models (qwen-turbo, auto-fast) for immediate replies. Use 客服消息 API for longer responses that need streaming.

小程序 Mini Program example

WeChat Mini Program calls your cloud function → cloud function calls AIPower → stream tokens back via WebSocket or 云开发 realtime.

// WeChat Cloud Function (Node.js)
const OpenAI = require("openai");
const cloud = require("wx-server-sdk");
cloud.init();

const aipower = new OpenAI({
  baseURL: "https://api.aipower.me/v1",
  apiKey: process.env.AIPOWER_API_KEY,
});

exports.main = async (event, context) => {
  const { userId, question, model = "auto" } = event;

  // Pick model based on user tier / question type
  const modelId = {
    auto: "deepseek/deepseek-chat",   // default — $0.34/M, great Chinese
    premium: "anthropic/claude-sonnet", // VIP users — $3.45/M
    free: "zhipu/glm-4-flash",         // nearly free — $0.01/M
  }[model] || "deepseek/deepseek-chat";

  const res = await aipower.chat.completions.create({
    model: modelId,
    messages: [
      { role: "system", content: "你是小程序内的 AI 助手，回答中文，结构清晰。" },
      { role: "user", content: question },
    ],
    user: userId,  // for per-user billing analytics
  });

  return {
    reply: res.choices[0].message.content,
    tokens: res.usage.total_tokens,
    model: modelId,
  };
};

Which model for Chinese users

Why Chinese developers use AIPower

WeChat-specific gotchas 微信平台坑点

FAQ

Building a different kind of bot?