Xiaomi miclaw product information and global closed beta test

Xiaomi miclaw marks our initial foray into the world of AI agents. Built upon Xiaomi's foundation model, Xiaomi MiMo, this interactive AI test product today begins closed beta testing. Designed to validate LLM execution within the "Human x Car x Home" smart ecosystem, Xiaomi miclaw explores the evolution of AI from mere conversational interfaces to a means of autonomous system-level execution. Its intelligence is anchored around four pivotal layers: kernel-level integration, personal context awareness, ecosystem connectivity, and continuous self-evolution.

It transforms the smartphone into a proactive AI tool. Upon understanding the user's intent and having received authorization, it orchestrates first-party apps, ecosystem services, and system tools to execute complex commands autonomously. It can also recognize and remember the user, so that ambiguous prompts can be deciphered to complete tasks accurately. Through a memory system, it refines its behavior and expands its capabilities over time, becoming more intuitive the more it is used.

In this cutting-edge test project, we are continuously optimizing stability, power consumption, and execution success rates in complex scenarios. Therefore, fluctuations in performance or intermittent errors may occur during high-complexity tasks. This closed beta is invite-only and is not open for public use at this stage. We do not recommend installing this on your daily driver. We advise tech and AI enthusiasts to conduct a full backup of all their data and only conduct testing within a controlled environment.

01 Foundational System-level Capabilities: Giving AI the Ability to Act

Why System Identity Is Essential

Xiaomi miclaw runs as a system-level application with core system capabilities.

Regardless of how powerful a model may be, AI running within a standard application sandbox is largely limited to answering questions or performing web searches. A system-level AI, however, operates differently. It can provide direct system-level control and execution capabilities.

50+ System-level Tools

Xiaomi miclaw packages the smartphone's core capabilities into more than 50 system-level tools and ecosystem services (with more constantly being added), enabling the AI to interact with and operate the device at a deeper level.

Each tool accepts structured parameters and returns execution results. The availability of specific tools depends on the permissions granted by the user.

Large Model–driven Reasoning and Execution Engine

Having tools alone is not enough. The key is getting the model to decide how and when to use them.

Xiaomi miclaw's core engine is built around a reasoning–execution loop:

User input → Model reasoning (selects tools and defines parameters) → Tool execution → Result feedback → Continued model reasoning → … → Task completion and response

At every step, the model independently determines which tool to call upon, which parameters to transmit, and whether the task has been completed. The execution of each tool is protected by an independent timeout mechanism, and the entire system operates on an asynchronous architecture that avoids blocking system threads.

Within its underlying layer, Xiaomi miclaw provides a unified interface for mainstream large-model protocols, allowing users to switch between different models without modifying the upper-layer logic. For users, the experience is streamlined and transparent. That is, they can see which tools the AI is calling upon and which step is currently being executed.

Context Management: Maintaining Coherence in Multiple Rounds of Conversations

The original objective can easily be lost in long conversations, particularly when multiple tools are involved. To address this, Xiaomi miclaw adopts a three-tier intelligent memory management system. It automatically preserves key decision points, dynamically compresses redundant interactions, and locally caches core instructions for efficiency. Even when executing 20 consecutive steps in complex workflows, the AI can still retain the original intent without requiring the user to repeat the context, ensuring a smooth and coherent experience throughout the entire task.

• Round Compression

Old conversations are compressed as a whole, while recent interactions are prioritized.

• Token Compression

When approaching the context window limit, messages are intelligently compressed at the message level.

At the same time, the system prompt architecture adopts a multi-layer design, with dynamic information injected into the first user message. This approach allows the prompt part to be cached by the model, thereby significantly reducing token usage. Internal testing shows that this method can reduce token overhead by 50–90%.

02 Personal Context: Helping AI to Truly Understand the User

From "What You Said" to "How You Live"

Most AI assistants lack sufficient contextual awareness and primarily focus on the message that a user sends at a given moment. After gaining authorization from the user, Xiaomi miclaw goes further, leveraging insights from usage patterns and personal context to provide more relevant and proactive assistance.

"Perception → Association → Judgment → Action"

Traditional applications typically adopt the following workflow: Perception → Action. For example, when a notification is received, the system simply displays it. Xiaomi miclaw introduces two additional steps in between: association and judgment. This forms a more intelligent workflow:

Perception → Association → Judgment → Action.

This expanded decision chain represents a key distinction between an AI agent and traditional software.

Example Use Case: Automated Trip Preparation

Imagine receiving a push notification: "Ticket purchased successfully: Flight MI123 Munich → Barcelona, Friday 10:00." What happens next?

Xiaomi miclaw executes 7 automated steps by chaining 6 different tools. While secure access to SMS and Calendar data is the prerequisite, the true effectiveness hinges on the AI agent's reasoning capabilities, powered by deep personal context.

Data Security

All conversation history, user profiles, and skill files are stored locally on the device. Cloud transmission occurs only for the current dialogue during inference, with all data being encrypted in transit. Once processed, no data is retained in the cloud, ensuring total data security and privacy. For tools performing sensitive operations (such as sending messages or creating calendar events), a confirmation prompt will appear before execution.

03 Ecosystem Connectivity: From a Single Device to a Connected Smart Life

When AI Can Control Home Devices

The first two chapters focused on the phone itself. The third layer extends AI beyond the phone and integrates it into the Xiaomi ecosystem.

Xiaomi Home app connects over 1 billion devices. Xiaomi miclaw implements a full Xiaomi Home protocol client, which can read device status and send control commands. With user authorization, it can theoretically control all IoT devices connected to the user's Xiaomi Home app.

The key role of this protocol client is translation. IoT device capabilities are usually defined in machine-readable specifications (parameter types, value ranges, constraints). Xiaomi miclaw compiles these specifications into natural language parameter descriptions that the large model can process. For example, the model receives instructions, such as: "Turn on the living room air conditioner; parameters: switch (on/off)."

Use Case: Automated Home Office Mode

Calendar event: "Important client meeting at 10:00" → AI determines that a quiet environment is needed:

• Phone is automatically silenced

• Robot vacuum is paused

• Incoming calls are handled based on urgency: calls from the boss trigger vibration alerts, while delivery personnel calls are logged silently

• After the meeting, all devices return to their former state, and missed calls and message summaries are sent

Traditional smart home systems rely on pre-set rules. Xiaomi miclaw is different: the AI makes real-time decisions based on context. For the same "meeting scheduled", the AI might only silence the phone if it's an internal weekly meeting, whereas if it's an important client meeting, it might coordinate all devices in the home.

The AI bases its judgment on calendar labels (e.g., "important client" vs. "weekly meeting"), past user habits during meetings, and the current status of the connected devices.

Open Protocol Extensions

The ecosystem is not limited to proprietary Xiaomi devices. There are also two ways to enable third-party integration:

• MCP (Model Context Protocol): An open standard for AI tool integration. Xiaomi miclaw implements a full MCP client, supporting standardized communication protocols, streaming responses, and asynchronous service discovery. Thousands of MCP tools already available on PCs could theoretically be accessed by the mobile AI.

• Open SDK: Xiaomi miclaw provides a third-party app integration SDK with signature-level permission protection. Third-party apps declare the capabilities they provide, and Xiaomi miclaw dynamically discovers and invokes them. In this model, apps tell the AI what they can do, and the AI calls upon them as needed.

04 Self-Evolution: Continuous Enhancement of AI Capabilities

The above three layers are the capabilities that Xiaomi miclaw launches with. But there is another point worth noting: what it can do is not fixed at the time of its release.

The capabilities of Xiaomi miclaw can simply be described as "tools to create tools":

• File-level Memory → Creating Arbitrary Data Structures

The AI uses this capability to autonomously design its memory system, creating memory directories, defining storage formats, and writing retrieval rules. The more it is used, the better it understands the user's preferences, frequent contacts, and historical tasks.

• Sub-agent Creation → Specialized Roles

Each sub-agent operates with a dedicated system prompt and whitelisted tools, working independently within a permission-controlled sandbox. For example, the user can have one sub-agent dedicated to managing schedules and another for summarizing information.

• MCP Service Configuration → Dynamic Access to External Capabilities

The AI can expand the range of callable tools at any time by integrating new MCP services.

• Script Execution → Sandbox Execution of Python/JavaScript

Provides general-purpose programming capabilities within a secure sandbox.

These four capabilities combined give Xiaomi miclaw a growth path: first, adjust its own behavior (modifying prompts and files), then expand its capabilities (creating sub-agents, integrating MCP), and finally consolidate its experience through its memory system—the more it is used, the better it understands the user.

No one predefines these workflows. The AI takes the basic tools and autonomously combines them into higher-level uses.

Closing Remarks

System-level foundations enable the AI to act. Personal context allows it to understand the user. Ecosystem connectivity lets it operate beyond the phone. Self-evolution ensures these capabilities continuously improve.

When large models are fully integrated into the mobile system, the phone itself becomes part of the AI's body.

Use-case Demonstration

Note:

*The following videos are based on an internal test build, recorded in Demo mode, and have been sped up and edited for demonstration purposes. Actual UI may vary; please refer to actual use.

*Availability of certain devices shown is subject to regional restrictions; actual user experience and scenarios may vary. Please refer to actual products, features, and functionality.

*All of the below content is generated by AI; please refer to actual use. Internet connection required. Check responses for accuracy.

Welcome Home

• Prompt: "I will bring my friend Mike home in half an hour. Please prepare the house to give Mike a warm welcome."

• Response:

▶ Videó megnyitása

Checking Supplies

• Prompt: "I'm at a Xiaomi Store. Please check what supplies I'm running low on at home."

• Response:

▶ Videó megnyitása

Schedule Management

• Prompt: "My schedule for March 21 (the day after tomorrow) is a bit messy, could you help me reorganize it?"

• Response:

▶ Videó megnyitása

Wake-up Call

• Prompt: "Wake me up at 7am tomorrow. If I don't get up, go all out to make me."

• Response:

▶ Videó megnyitása

Daily Briefing

• Prompt: "Brief me on the day's schedule, then search for 3–4 AI and finance-related news items. Feel free to also include other content. Repeat this every morning at 8:16am in audio format."

• Response:

▶ Videó megnyitása

Financial Radar

• Prompt: "Can you monitor the spot gold price for me? Provide an audio update every two minutes."

• Response:

▶ Videó megnyitása