WeChat's AI has finally moved.
On the very same day as Apple's WWDC, WeChat did something potentially more significant than Apple, releasing an unassuming announcement: "Guidance for Developers to Access the WeChat AI Ecosystem."
Starting today, Mini Program developers can grant authorization, allowing WeChat AI to read, operate, and invoke the functionalities of Mini Programs.
WeChat offers two access modes. The first is "Automatic Mode," with a nearly zero barrier to entry. Developers simply toggle a switch, and the platform itself reads the source code, analyzes pages, figures out what the Mini Program can do, and then the AI can directly operate it, all without writing a single line of code.
The other mode is called "Developer Mode," where developers create customized Skills, which are called by the AI after passing review. Both modes can be enabled simultaneously. Meituan has already announced its integration.
This should not be understood merely as the launch of another new feature. Instead, it signals that WeChat is turning its entire ecosystem—millions of Mini Programs, WeChat Pay, service notifications, Official Accounts—into the execution layer for AI.
Examining the Skill Documentation: How WeChat AI Invokes Mini Programs
The WeChat open documentation publicly shares the technical specifications for Mini Programs to integrate AI Skills. A close look reveals many design details hidden within.
Official skill documentation guide 👇🏻:
https://developers.weixin.qq.com/miniprogram/dev/ai/best-practices.html
From an architectural perspective, those familiar with AI development will immediately recognize it: it's essentially MCP (Model Context Protocol). The `mcp.json` declares the function and parameters of each atomic interface, and `SKILL.md` describes how the entire business process runs. This is almost identical to the MCP+Skills architecture found in Claude, Cursor, or VS Code. WeChat didn't reinvent the wheel; it directly adopted the industry-standard that is currently converging.
In the guidance, WeChat provides a clear "attention weight" system. When the AI decides which interface to call and what parameters to generate, it gives the highest priority to the content returned by the interface (five stars), followed by the interface description (four stars) and parameter descriptions (four stars) in `mcp.json`. `SKILL.md` ranks last (three stars). This means where developers write something matters more than what they write—the weight the AI gives to the same rule is completely different if it's written in the interface return versus in `SKILL.md`.
At the interface return level, there is a core specification: a two-stage "Fact + Action" format. First, tell the AI "what happened," then tell it "what to do next." If only the action is written without the fact, the AI might interpret "display card" as "prepare to call the next interface" and skip user confirmation. This is a rule learned after stepping into many pitfalls.
Fourth, parameter passing prioritizes using IDs over natural language. Taking the "Coffee Ordering" scenario in the diagram as an example, after the user states a need, the AI understands the vague intent and handles selection, modification, specification changes, and payment processing, all without leaving the chat dialog.
This design reveals a signal: WeChat has already run enough practical cases, knows where the pitfalls lie in AI calling external services, and has solidified these experiences into developer norms.
In fact, comparing WeChat Mini Programs, which are also known for their "ecosystem," to Apple's apps, WeChat possesses a kind of "God's-eye view" over its own ecosystem. This is the prerequisite for all this implementation.
Why It Might Be More Important Than Apple's AI
This year at WWDC, Apple released the new Siri AI. Despite having Google Gemini integrated at the underlying level and supporting natural language creation for Shortcuts, it didn't spark much discussion.
A closer look reveals the gap: Apple is making AI coordinate some native functions within the iOS system. Once it involves third-party applications—those apps installed on your phone—it becomes strained.
Take Ele.me, for example. Its code runs on Ele.me's own servers, which Apple cannot read. For Siri to call Ele.me, Ele.me's engineers must proactively connect to the App Intents API, negotiating and integrating one by one, a time-consuming and labor-intensive process.
What WeChat is doing is enabling AI to directly operate millions of third-party services, because Mini Programs are different. The code for every Mini Program, from developer submission, through WeChat's review, to finally running on the user's phone, remains entirely within WeChat's technical system throughout the entire process. During the review phase, WeChat can scan the code, automatically analyzing "what pages this Mini Program has, what it can do, what its inputs and outputs are."
This is why "Automatic Mode" is possible—developers don't need to write a single line of code. They just flip a switch, and WeChat itself can translate your Mini Program into a tool the AI can invoke. WeChat's foundational architecture naturally supports this. It possesses a "God's-eye view," enabling scheduling based on centralization.
Apple does not have this architectural advantage, and neither does Google.
Also noteworthy is the recent rumor that WeChat is collaborating with Huawei, Honor, Xiaomi, OPPO, and vivo to launch A2A (Agent-to-Agent) assistant capabilities, allowing users to directly initiate WeChat audio/video calls or send messages via their phone's voice assistant.
Internally, WeChat AI can invoke millions of Mini Programs. Externally, smartphone manufacturers' AI assistants can invoke WeChat. WeChat is becoming the super connector of the AI era, a service hub that all AIs can access.
The Old Prophecy of "WeChat OS"
When Mini Programs were launched, many joked that WeChat wanted to become "WeChat OS." Back then, it was more of a figure of speech—Mini Programs replaced some app functionalities but were essentially a "light application platform."
More coincidentally, the centralized review mechanism designed at the time was for quality and security control. Nine years later, this design, initially criticized as "excessive control," has unexpectedly become an infrastructure advantage in the AI era. The distributed App ecosystem (Apple/Android) seemed more "free" at the time, but now it has become an obstacle to AI integration.
An old prophecy, due to the emergence of new-era technology—AI—has taken on a transformative change.
Previously, when writing about OpenClaw and Feishu, I mentioned a judgment: IM (Instant Messaging) is the most natural entry point for AI Agents because dialogue itself is the most natural interaction between humans and AI, and the service ecosystem (bots, payments, mini-programs) inherent to IM allows AI not only to "chat" but also to "do." Feishu is already moving in this direction, launching enhanced Bot APIs and AI Agent nodes.
However, Feishu is an enterprise collaboration tool, covering work scenarios. WeChat has a distinctly different breadth—1.432 billion monthly active users, hundreds of Mini Programs across niche sectors, covering almost all service needs of a person's daily life, from ordering takeout to hospital registration to buying flight tickets to paying utility bills.
If WeChat AI can indeed smoothly invoke these Mini Programs to complete tasks, then, as the prophecy said, it becomes an operating system operated with natural language.
A user says, "Help me book a high-speed rail ticket from Beijing to Shanghai for tomorrow at 3 pm." The AI deconstructs the intent, invokes the 12306 Mini Program to check tickets, select seats, completes the order via WeChat Pay, all without leaving WeChat. This path can theoretically be executed today.
Of course, there is distance between theory and reality. AI invocation involving payment scenarios requires a near-zero error tolerance—ordering the wrong coffee is a minor issue, but buying the wrong flight ticket is a major one. The accuracy requirements for the underlying model are far higher than for conversational scenarios. This is also a common bottleneck facing AI Agent implementation globally: the gap between "able to chat" and "able to get things done" is not measured by technical metrics but by trust.
But WeChat has at least done one thing right: it didn't build a service network from scratch. Over the years, what ChatGPT has been doing is first having a smart brain, then connecting one by one to Shopify, DoorDash, Stripe, each connection built from the ground up. To this day, transaction-related queries still account for less than 3%.
The real change that is about to happen might be imperceptible to most users. One day, you type into WeChat, "Help me book a ticket to Shanghai for 9 pm tonight," and then it's done. You might not even know which Mini Program was called in the background or what payment process was followed.
This "imperceptible completion" is the true mark of a mature AI Agent. WeChat is closer to this step than anyone else.
This article is from the WeChat Official Account "APPSO", author: APPSO discovering tomorrow's products
