WeChat's AI has finally moved.
On the same day as Apple’s WWDC, WeChat did something potentially more significant: it released a simple, unadorned announcement titled “Guidelines for Developers Integrating with WeChat’s AI Ecosystem.”
Starting today,小程序 developers can grant authorization for WeChat AI to read, operate, and invoke小程序 functions.
WeChat offers two integration methods: one is "Automatic Mode," with virtually no barriers to entry—developers simply enable a switch, and the platform automatically reads the source code, analyzes the page, understands what the mini-program can do, and then the AI can immediately begin operating without writing a single line of code.
Another mode is called "Development Mode," where developers create custom skills themselves and, after approval, these skills can be invoked by AI. Both modes can be enabled simultaneously. Meituan has already announced its integration.
This shouldn't just be seen as the launch of another new feature; rather, it shows that WeChat is transforming its entire ecosystem—millions of mini-programs, WeChat Pay, service notifications, and official accounts—into an execution layer for AI.
Look into the Skill documentation to see how WeChat AI calls mini-programs.
The WeChat open documentation publicly outlines the technical specifications for integrating AI skills into mini programs; upon closer inspection, it contains many design details.
Official skill documentation: 👇🏻
https://developers.weixin.qq.com/miniprogram/dev/ai/best-practices.html
From an architectural perspective, anyone with AI development experience will immediately recognize it as essentially MCP. The mcp.json file declares the functionality and parameters of each atomic interface, while the SKILL.md file describes how the entire business process flows—this is nearly identical to the MCP + Skills architecture used in Claude, Cursor, and VS Code. WeChat did not create a new system from scratch but instead adopted the industry-standard approach that is currently converging.
In the guidance framework, WeChat provides a clear system of "attention weights." When deciding which API to call and what parameters to generate, the AI prioritizes the content returned by the API (five stars), followed by the API description and parameter descriptions in mcp.json (four stars each), with SKILL.md ranked last (three stars). This means that where developers place information is more important than what they write—identical rules receive entirely different weights depending on whether they are written in the API response or in SKILL.md.
At the API response level, there is one core guideline: a two-part structure of “Fact + Action.” First, tell the AI “what happened,” then tell it “what to do next.” If you only state the action without the fact, the AI might interpret “display card” as “prepare to call the next API” and skip user confirmation. This is a rule learned only after encountering many pitfalls.
Fourth, prefer using IDs over natural language for parameter passing. Using the "coffee order" scenario in the diagram as an example, after the user submits their request, the AI understands ambiguous intent, selects options, modifies specifications, and processes payment—all without leaving the chat window.
This design signals that WeChat has already tested numerous real-world scenarios and understands the pitfalls of AI calling external services, having codified these insights into developer guidelines.
In fact, when comparing WeChat Mini Programs with Apple apps—both known for their ecosystems—WeChat has a "god's-eye view" of its own ecosystem, which is the foundation for everything it achieves.
How is it more important than Apple's AI?
This year, Apple's new Siri AI released at WWDC, despite being powered by Google Gemini underneath and supporting natural language creation for Shortcuts, did not generate much discussion.
Look closer, and you’ll see the difference: Apple’s approach is to have AI coordinate certain native functions within the iOS system, but it struggles when it comes to third-party apps—those installed on your phone.
For example, Ele.me runs its code on its own servers, which Apple cannot access. For Siri to interact with Ele.me, Ele.me’s engineers must proactively integrate with the App Intents API—one by one, through individual negotiations and implementations, which is time-consuming and labor-intensive.
WeChat enables AI to directly operate millions of third-party services because mini-programs are different. The code for each mini-program—from submission by the developer, through WeChat’s review process, to final execution on the user’s device—remains entirely within WeChat’s technical ecosystem. During the review phase, WeChat can scan the code and automatically analyze: “What pages does this mini-program have? What functions can it perform? What are its inputs and outputs?”
Therefore, the "automatic mode" can function—developers don’t need to write a single line of code; simply flip a switch, and WeChat will automatically translate your mini-program into a tool that AI can invoke. WeChat’s underlying architecture natively supports this, as it has a “god’s-eye view” and can orchestrate operations through centralization.
This architectural advantage is not possessed by Apple or Google.
It is also worth noting that recently it was reported that WeChat is collaborating with Huawei, Honor, Xiaomi, OPPO, and vivo to introduce Agent-to-Agent (A2A) assistant capabilities, allowing users to initiate WeChat audio or video calls or send messages directly through their phone’s voice assistant.
Internally, WeChat AI can access millions of mini-programs; externally, smartphone manufacturers' AI assistants can integrate with WeChat. WeChat is becoming the super connector of the AI era—a central service hub that enables all AI systems to connect.
The old prophecy of "WeChat OS"
When the mini-program was launched, many jokingly said WeChat was building a "WeChat OS." At the time, this was more of a rhetorical statement—mini-programs replaced some functions of apps, but were still essentially a "lightweight application platform."
Ironically, the centralized review system designed at the time was intended to control quality and security. But nine years later, this design, once criticized as "overly controlled," has unexpectedly become an infrastructure advantage in the AI era. The decentralized app ecosystem (Apple/Android) seemed more "free" back then, but now it has become an obstacle to AI integration.
An old prophecy has been dramatically transformed by the emergence of new-era technology—AI.
When I previously wrote about OpenClaw and Feishu, I made the observation that IM is the most natural entry point for AI agents, because conversation is the most natural form of human-AI interaction, and IM’s built-in service ecosystem (bots, payments, mini-programs) enables AI to not just “chat” but also “do.” Feishu is already moving in this direction, having launched enhanced Bot APIs and AI Agent nodes.
However, Feishu is an enterprise collaboration tool focused on office scenarios. WeChat, by contrast, has a vastly different scope—1.432 billion monthly active users and hundreds of specialized mini-programs covering everything from ordering food and booking doctor’s appointments to buying plane tickets and paying utility bills, essentially encompassing all the service needs of daily life.
If WeChat AI can truly seamlessly invoke these mini-programs to complete tasks, then, as predicted, it has become an operating system controlled by natural language.
The user says, “Book me a high-speed train from Beijing to Shanghai for tomorrow at 3 p.m.” The AI parses the intent, invokes the 12306 mini-program to check availability, select a seat, and complete the booking via WeChat Pay—all within WeChat. This entire workflow could theoretically be executed today.
Of course, there is still a gap between theory and reality. AI interactions involving payment scenarios have near-zero tolerance for error—accidentally ordering a cup of coffee is minor, but mistakenly purchasing a plane ticket is major. The accuracy requirements for underlying models far exceed those of conversational scenarios. This is a common bottleneck facing global AI agents: the transition from “able to chat” to “able to act” is not bridged by technical metrics, but by trust.
But WeChat got at least one thing right: it didn’t build its service network from scratch. Over the years, ChatGPT has been doing the opposite—starting with a smart brain and then connecting one by one to platforms like Shopify, DoorDash, and Stripe, each requiring a brand-new integration; even today, transactions account for less than 3% of queries.
The real changes to come may happen silently for most users. One day, you type in WeChat, “Book me a ticket to Shanghai for 9 p.m. tonight,” and it’s done—you never even realize which mini-program was invoked or what payment process was used.
This kind of "seamless completion" is the true hallmark of a mature AI agent, and WeChat is closer to this step than anyone else.
This article is from the WeChat public account "APPSO," authored by APPSO, discovering tomorrow's products.