How to Use AutoGLM to Automate Phone Tasks on Android
AutoGLM is a mobile AI agent from Zhipu AI that controls a real Android phone through screenshots and taps. You give it a goal like “order a coffee on Meituan” and it opens the app, finds the buttons, and finishes the flow. Setup takes about 15 minutes on a spare device.
What Is AutoGLM?
AutoGLM reads the screen, decides the next tap, and executes it. Unlike a chat assistant, it acts inside apps you already have installed. Zhipu shipped it as a research preview tied to the GLM model family, and the Android build runs as a separate app that needs accessibility permissions.
It is not a script recorder. There is no fixed macro. The agent re-reads the UI on every step, which is why it can handle pop-ups and changed layouts that break old automation tools.
What Do You Need Before You Start?
- An Android phone or emulator running Android 10 or newer
- A Zhipu AI account and API access to the GLM model
- Accessibility and screen-capture permissions granted to the AutoGLM app
- A spare device, or an emulator. Do not hand it your main phone on day one.
An emulator is the safer path. You can watch exactly what the agent does and kill it if it starts tapping the wrong things.
How Do You Install and Connect AutoGLM?
- Download the AutoGLM Android build from the Zhipu release page and install the APK.
- Open the app and paste your GLM API key in settings.
- Grant Accessibility, Screen Capture, and Overlay permissions.
- Run the built-in test task and confirm the agent can open Settings.
If the test task stalls, the usual cause is a missing overlay permission. The agent needs to draw its control bubble on top of other apps to report progress.
Is AutoGLM Free?
No. The app itself is free to download, but every action costs tokens against your Zhipu API balance. A simple task like opening an app and reading a screen uses a handful of calls. A long multi-app flow can burn through far more, because each screenshot plus reasoning pass is a billed request. Budget roughly a few cents per short task on standard pricing, and test on cheap tasks first.
What Tasks Does It Handle Well?
Short, single-app flows are where it shines. Reading order status, filling a delivery form, or navigating to a setting menu all work reliably. Multi-app chains with logins and CAPTCHAs are where it struggles. If a flow needs a SMS code typed in, the agent will pause and wait for you.
For coding and desktop work, AutoGLM is the wrong tool. Pair it with a desktop agent instead – see our GLM-5 guide for the text-model side of the same family.
Where Does It Fail?
Two real limits. First, latency. Each step takes seconds, so a 20-tap flow can run for a minute or more. Second, payments and irreversible actions. The agent will happily walk toward a confirm button, and there is no built-in budget cap. Keep it away from anything that spends money until you trust the behavior.
Quick Summary
- AutoGLM controls a real Android phone through screenshots and taps, not scripts.
- You need a GLM API key; every action is billed per request.
- Best for short single-app tasks; weak on multi-app flows with logins.
- Use an emulator and disable payment apps until you trust it.
- Expect seconds of latency per step, not instant taps.
