Migpt Xiaomi Assistant

MiGPT Xiaomi Voice Assistant Deployment

Deploy mi-gpt v4.2.0+ on a Xiaomi smart speaker to replace the built-in AI with a custom LLM assistant.

Prerequisites

• Node.js >= 18
• Xiaomi account with a supported smart speaker
• An OpenAI-compatible LLM API (DashScope, DeepSeek, OpenRouter, etc.)

Quick Start

1. Initialize project

mkdir -p ~/projects/mi-gpt && cd ~/projects/mi-gpt

Create package.json:

{
  "type": "module",
  "dependencies": { "dotenv": "^17.0.0", "mi-gpt": "^4.2.0" }
}

npm install

2. Create files

app.js:

import "dotenv/config";
import config from "./.migpt.js";
import { MiGPT } from "mi-gpt";
const client = MiGPT.create(config);
await client.start();

.env:

OPENAI_MODEL=your-model
OPENAI_API_KEY=sk-xxx
OPENAI_BASE_URL=https://your-api-endpoint/v1

.migpt.js — See references/config-template.md for full template with comments.

3. Find your device's MIoT commands

Different speaker models use different MIoT siid/aiid values. Look up your model in mi-gpt issues or the MIoT spec.

Common values:

4. Run

nohup node app.js > migpt.log 2>&1 &

Critical Issues & Fixes

Issue 1: MiNA TTS returns success but no audio (X08E and some models)

Symptom: mina.play({tts: '...'}) returns true but speaker stays silent.

Cause: Some models (notably X08E) don't support MiNA's ubus mibrain/text_to_speech. They require MIoT's doAction.

Fix: Ensure ttsCommand is set correctly. MiGPT will use MIoT doAction(siid, aiid, text) when ttsCommand is configured.

Issue 2: MIoT login triggers security verification every time

Symptom: securityStatus: 16 on every startup, even after completing SMS verification multiple times.

Root cause: mi-service-lite generates a random deviceId on every startup → Xiaomi treats each login as a new device → perpetual verification loop.

Fix: See references/miot-auth-bypass.md for the browser cookie injection method.

Issue 3: streamResponse causes hang on some models

Symptom: MiGPT says "让我想想" then freezes.

Cause: streamResponse: true polls MiNA.getStatus() for playback state. On models where MiNA doesn't control the speaker (e.g., X08E), status always returns "paused", causing the poll loop to malfunction.

Fix: Set streamResponse: false. If you also need keepAlive (continuous conversation) mode, patch enterKeepAlive() in node_modules/mi-gpt/dist/index.js to remove the streamResponse check. See references/patches.md.

Issue 4: Native 小爱 AI races with custom AI

Symptom: Both 小爱 and your custom AI respond to the same query ("split personality").

Mitigations (pick one or combine):

1. Set onAIAsking: ["嗯"] — immediately interrupts 小爱 with a short TTS
2. Disable voice reply in 小爱 App (device settings → 语音回复 → off)
3. Use 全量接管模式: callAIKeywords: [""] — responds to all queries

Issue 5: Wake word misrecognition

Symptom: 小爱's ASR mishears custom keywords (e.g., "花卷卷" → "花姐姐").

Fix: Add common misrecognition variants to callAIKeywords and wakeUpKeywords.

LLM Model Selection

For voice assistants, latency > intelligence. Test your API endpoint:

time curl -s --max-time 15 YOUR_BASE_URL/chat/completions \
  -H "Authorization: Bearer YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"MODEL","messages":[{"role":"user","content":"一句话回答：今天星期几"}],"max_tokens":30}'

Target: < 3 seconds for acceptable voice experience. Total end-to-end latency ≈ LLM time + 3-4s (polling + MIoT overhead).

Reference Files

• references/config-template.md — Full .migpt.js configuration template with all options explained
• references/miot-auth-bypass.md — Browser cookie injection to bypass MIoT security verification
• references/patches.md — Required code patches for mi-gpt and mi-service-lite
• references/latency-analysis.md — End-to-end latency breakdown and optimization tips