Xiaomi Miclaw: Xiaomi's AI Agent Built on MiMo Model Enters Closed Beta

Forum topic · 小凯 · 2026-03-06

Summary

Xiaomi has announced Xiaomi Miclaw, an AI Agent product built on the MiMo large language model, which began small-scale closed beta testing on March 6, 2026. Miclaw is designed to validate large-model execution capabilities across Xiaomi's 'Human x Car x Home' ecosystem, moving beyond conversational AI toward system-level action. The product features a four-layer architecture: (1) system-level capabilities with 50+ native tools, non-sandboxed system app access, an autonomous reasoning-execution engine, three-tier context memory supporting 20-step operations, and prompt caching that cuts token costs by 50–90%; (2) personal context understanding with on-device conversation storage, enabling scenarios like automatic travel preparation and duplicate-subscription detection; (3) ecosystem connectivity, including a full Mijia protocol client controlling over 1 billion IoT devices, MCP client support, and an open SDK; (4) self-evolution via file-level memory, sub-agent creation, and sandboxed Python/JavaScript execution. Testing is invitation-only and not open to public registration.

Xiaomi has unveiled Xiaomi Miclaw, an AI Agent product built on the company's MiMo large language model. Small-scale closed beta testing began on March 6, 2026.

Product Positioning

Miclaw focuses on validating large-model execution capabilities within Xiaomi's "Human x Car x Home" ecosystem, exploring how a model can move from "conversational ability" to "system-level execution ability."

Four-Layer Capability Architecture

1. System-Level Capabilities: Giving AI Hands and Feet

System app identity: Access to core system capabilities, not running in a sandbox
50+ system-level tools: Wrapping phone system capabilities and ecosystem services
Reasoning-execution engine: The model autonomously decides tool call order, with streaming progress updates
Context management: Three-tier intelligent memory management; no context loss across 20-step complex operations
Token optimization: Multi-level prompt caching design saves 50%–90% of token costs

2. Personal Context: Truly Understanding the User

Perceive → associate → judge → act: The dividing line between AI Agents and traditional software
Example — automatic travel preparation:
Example — spending insights:
Data security: Conversation history stored locally; only the current conversation is sent to the cloud, with encrypted transmission

3. Ecosystem Connectivity: From One Phone to the Whole Smart Life

Mijia ecosystem: Full Mijia protocol client, able to control over 1 billion devices
Capability translation: Compiling IoT device specifications into natural language the LLM can understand
Example — home office mode:
MCP protocol: Full MCP client, supporting thousands of MCP tools on PC
Open SDK: Third-party apps can proactively declare capabilities for the AI to invoke on demand

4. Self-Evolution: Continuous Capability Growth

File-level memory: Creating arbitrary data structures and designing its own memory system
Sub-agent creation: Specialized division of labor with independent prompts and tool whitelists
MCP service configuration: Dynamically connecting external capabilities
Script execution: Running Python/JavaScript in a sandbox

Scenario Demonstrations

"Bringing a friend home in half an hour" → automatically prepares a welcome scene (lights / AC / music)
"Look at food photos in my album and help me lose 3kg next week" → analyzes eating habits and creates a plan
"Wake me at 7 a.m. tomorrow — if I won't get up, annoy me hard"
"Every morning at 8:16, announce my schedule plus AI finance news by voice"
"Report gold prices every two minutes"

Testing Information

Closed beta, invitation-only; no public recruitment
Stability, power consumption, and complex-scenario success rates are still being optimized
Upgrading your primary device is not recommended; back up your data first

Original source: https://mp.weixin.qq.com/s/TLU6WXkgI-7Ph2ebGQKARg