WeChat DevTools + AI: I never left the window from writing code to deploying

Last week I poked around WeChat DevTools and found it had quietly changed.

There's a new CodeBuddy in the extension marketplace. Installing it adds an AI chat window to the right side of the IDE. I'd been wanting to try writing a mini program directly in WeChat DevTools with AI, and deploying to CloudBase, without ever leaving the window.

I ran through it once and logged the real experience. It works, but there are pitfalls.

First, get three things straight​

Before starting, three concepts need sorting out, or the steps ahead will be confusing.

CodeBuddy extension — WeChat DevTools 2.0 and above; search "CodeBuddy" in the extension marketplace to install. After installing, a new icon appears in the toolbar; click it for the AI chat window. It reads your code context, helps you write and modify code, and explains errors. Essentially an AI coding assistant embedded in WeChat DevTools.

CloudBase MCP — the bridge connecting AI and cloud development services. Once installed, AI doesn't just write code — it operates cloud resources directly: creating database tables, deploying cloud functions, configuring static hosting, all done in conversation. Installation is simple: CodeBuddy's settings has an MCP marketplace; search "CloudBase MCP" and install with one click.

CloudBase Skills — think of it as an "experience pack" for AI. How to use the mini program SDK, how to configure database permissions, what the best practices for cloud functions are — this domain knowledge that AI might not know, Skills packages up. Install command:

npx skills add tencentcloudbase/cloudbase-skills

Or just say "install CloudBase Skills" in the CodeBuddy chat and it runs the command for you.

In short, the three's relationship: CodeBuddy handles conversation and execution, MCP handles connection and permissions, Skills make AI knowledgeable.

Hands-on: build a to-do mini program​

Enough talk — run a real case.

I picked the most classic scenario — a to-do mini program. Simple features: add a to-do, mark complete, delete. Data stored in the CloudBase cloud database.

Environment prep​

Not much needed:

1. WeChat DevTools 2.0+ — get the latest stable from the official WeChat download page
2. Node.js v18.15+ — needed to install Skills
3. A CloudBase environment — new users get free quota; open it in the cloud development console

Steps to install the CodeBuddy extension:

• Open WeChat DevTools, use Shift + Command + X (Mac) or Ctrl + Shift + X (Windows) to open the extensions panel
• Search "CodeBuddy"
• Install "Tencent Cloud Code Assistant CodeBuddy"

After installing, click the CodeBuddy icon in the toolbar, enter settings, and install "CloudBase MCP" from the MCP marketplace.

At this point the MCP tool list should be visible at the bottom of the AI chat window. If the tool count shows 0, restart WeChat DevTools.

Let AI scaffold the project​

Open the CodeBuddy chat window and say:

Create a to-do mini program with features: add to-do, mark complete, delete to-do. Store data in the CloudBase cloud database. Use native WeChat mini program development, no framework.

AI starts outputting a few seconds later. It does these things:

1. Creates a todos collection in the cloud database
2. Generates the mini program page structure — home is a list page with an input box and the to-do list
3. Writes cloud functions for CRUD
4. Wires up the page event bindings and database operations

The output structure looks roughly like:

miniprogram/
├── pages/
│   └── index/
│       ├── index.wxml    # page structure
│       ├── index.wxss    # styles
│       ├── index.js      # logic
│       └── index.json    # config
├── app.js
├── app.json
└── app.wxss

cloudfunctions/
├── addTodo/
│   └── index.js
├── getTodos/
│   └── index.js
├── updateTodo/
│   └── index.js
└── deleteTodo/
    └── index.js

Honestly, the first version's completion was quite high. The basic CRUD logic was fine, and the page layout was reasonable.

Where you need to do it yourself​

It's not usable as-is, though. I changed a few things:

Style tweaks. AI's generated styles lean conservative — spacing, font sizes and colors are all "safe", but a bit plain. I adjusted the wxss myself, adding rounded corners and shadows so it doesn't look like a template.

Cloud function env ID. In AI-generated cloud function code, cloud.init() sometimes lacks the env ID and needs manual fixing:

wx.cloud.init({
  env: 'your-env-id'  // swap in your own env ID
})

Permission rules. AI-created database collections default to "creator read/write only", which is enough for a to-do app. If your scenario needs other permission config, you'll change it yourself.

Deploy​

Once the code is written, hand deployment to AI too.

Continue in the CodeBuddy chat:

Deploy the cloud functions to CloudBase, then preview the mini program for me

MCP handles these:

1. Deploy cloud functions — AI calls the cloud function deploy API via MCP, pushing the four functions addTodo, getTodos, updateTodo, deleteTodo to the cloud. No right-click menus, no manually selecting "upload and deploy"
2. Confirm the result — AI tells you each function's deploy status in chat; on success it continues, on failure it tries to diagnose

If you also want the front end hosted on CloudBase (say for an H5 version), tell AI "deploy the front end to CloudBase static hosting."

The mini program's own preview and upload still require clicking "Preview" and "Upload" in WeChat DevTools — that step is WeChat's platform flow, which MCP doesn't cover. But from writing code to cloud function deployment, it all happened in conversation, with no manual cloud-resource operations.

Through the whole flow, your hands only touched the mouse twice: installing the extension, and clicking preview/upload at the end. All the code generation, database creation and cloud function deployment in between was done by AI.

Pitfall log​

After a full run, the honest feelings.

What AI handles well:

• Basic CRUD logic — standard, clean
• Page layout — clear wxml structure, correct data binding
• Database design — sensible field naming and index suggestions
• Cloud function skeletons — basic functions run directly

What AI struggles with:

• Complex interaction animations — swipe-to-delete, pull-to-refresh feel: AI's solutions work but aren't refined
• WeChat-specific API parameter details — some API parameters change across versions, and AI may not keep up
• Fine style tuning — AI's styles are okay but far from "good-looking"; the design sense is still on you

MCP occasionally drops — hit one disconnect in about two hours; restarting WeChat DevTools fixed it. Not a big deal, but annoying if you're mid-way through a long AI generation task.

Skills coverage — currently covers core scenarios like database, cloud functions and auth. If you're doing business-y features like WeChat Pay, subscription messages or customer-service messages, Skills can't help much — you'll read the docs yourself.

Do the math​

Let's use data.

Time:

Is this comparison padded? A bit. The 5 traditional hours include doc-reading time; an expert might do it in 3 hours. But AI genuinely compressed the dreariest "scaffolding" phase from hours to minutes — that improvement is real.

Fits:

• ✅ Prototype validation, quick MVP — best fit
• ✅ Simple utility mini programs — mostly CRUD, AI covers 80% of the code
• ⚠️ Complex business-logic projects — AI scaffolds, but you fill in the business details
• ❌ Refactoring existing large projects — AI has limited context on big projects; better to start fresh

Learning cost:

If you already have mini program experience, the onboarding cost is near zero — just installing two extensions.

If you're new, I suggest skipping AI first and manually running the basic WeChat cloud development flow once. The WeChat official docs recommend the same: complete the basic flow manually once, then bring in MCP/Skill.

Two closing thoughts​

WeChat DevTools embedding AI coding capability is the right direction. For mini program developers, not jumping between tools is genuinely a relief.

But it's not the strongest AI coding tool right now. If you only look at AI coding smoothness, Cursor is better at the moment. WeChat DevTools' advantage is "one-stop" — mini program simulator, real-device debugging, upload and review all in one window.

Add the CloudBase CLI V3 released two days ago (details) — 15 top-level command modules designed for AI agents — and AI will be able to do even more.

Whether it fits your scenario, you'll know after running it once yourself.

Reference links:

• CloudBase AI Toolkit overview
• WeChat DevTools config guide
• WeChat official MCP-assisted development guide
• CloudBase CLI V3 details
• Video tutorials and cases
• CloudBase AI ToolKit open-source repo