Xiaomi Miclaw: Mobile AI Agent Enters Small‑Scale Closed Beta

Overview

Xiaomi Miclaw is an AI interaction testing product based on the MiMo large model, now entering a small‑scale closed test. It aims to verify the large model’s execution ability across Xiaomi’s "person‑car‑home" ecosystem, moving from pure dialogue to system‑level task execution.

System‑level Capability

Running with system‑application identity, Miclaw gains core system app privileges. It wraps more than 50 system‑level tools and ecosystem services, each accepting structured parameters and returning execution results, with availability determined by user‑granted permissions.

The core engine follows an inference‑execution loop: user input → model inference (select tool, set parameters) → tool execution → result returned → model continues inference → … until the task completes. The model decides at each step which tool to call, what parameters to pass, and whether the step is finished. Tool execution is protected by independent timeouts and runs asynchronously, avoiding blockage of system threads.

Underlying large‑model protocols are abstracted uniformly, allowing model swaps without changing upper‑level logic. The user experience updates in a streaming fashion, showing which tool the AI is invoking and its progress.

Context Management

Miclaw implements three‑level intelligent memory: round‑compression (compressing older dialogue turns), token‑compression (smartly trimming messages near the token limit), and dynamic prompt injection that saves 50‑90% of token overhead. This ensures long conversations retain essential context without losing the original intent.

Ecosystem Interconnection

Beyond the phone, Miclaw connects to Xiaomi’s IoT ecosystem via a full Mi Home protocol client. It translates device capability specifications into natural‑language parameter descriptions, enabling commands such as "turn on the living‑room AC, parameter: switch (on/off)". Example scenarios include automatically silencing the phone, pausing a robot vacuum, and adjusting device states based on calendar events.

Two open channels extend the ecosystem: the Model Context Protocol (MCP) client supports standardized communication, streaming responses, and asynchronous service discovery; an open SDK lets third‑party apps declare their tool capabilities, which Miclaw discovers and invokes dynamically.

Self‑Evolution

Miclaw’s meta‑capability is "a tool that creates tools". It supports file‑level memory (creating arbitrary data structures), sub‑agent creation (specialized agents with independent prompts and tool whitelists), MCP service configuration (dynamic external tool integration), and sandboxed script execution (Python/JavaScript) for general programming.

These layers form a growth path: adjusting behavior via prompts and files, expanding abilities by adding sub‑agents or MCP tools, and finally consolidating experience in a memory system that becomes more personalized with use.

Data Security

All conversation history, user settings, and skill files reside locally; only the current dialogue message is transmitted to the cloud during inference, encrypted in transit. Sensitive tools (e.g., sending SMS, creating calendar events) trigger a confirmation dialog before execution.

Demo Scenarios

Travel preparation: after receiving a ticket purchase notification, Miclaw orchestrates seven steps across six tools to arrange transport, accommodation, and reminders.

Consumption insight: reads three months of bank SMS, correlates overlapping video‑platform subscriptions, and suggests unsubscribing to save ~400 CNY per year.

Home office mode: detects a calendar meeting labeled "important client", silences the phone, pauses the robot vacuum, and adjusts notification priority based on context.

Self‑evolution showcase: creates sub‑agents for scheduling, information summarization, and runs sandboxed scripts to extend functionality.

These examples illustrate how Miclaw transforms the phone into an AI‑driven extension of the user’s digital life, gradually evolving its capabilities through continuous interaction.