Best Open Source Phone Agent Frameworks: Open-AutoGLM, Mobilerun, and mobile-use

Choose a Framework by the Job

The best open source phone-agent framework depends on what you are trying to build. A visual Android research agent, a traceable automation harness, and a structured mobile UI task runner are not the same project. Start with the work shape, then check the device bridge, model path, execution surface, license, and maintenance load.

This is a selection guide, not a benchmark. I am not ranking success rates or latency. The right choice is the one whose runtime, model inference, and phone execution path match the task you can actually maintain.

Compare Setup, Execution, Models, and Licensing

Separate four layers before choosing. The framework is the code you run. Model inference is where the LLM or vision model actually reasons. Phone execution is how taps, text input, screenshots, accessibility data, or device actions reach the phone. Inspection is how you debug failures after the task goes wrong.

Open source does not make the full system free. You may still pay for model API calls, hosted inference, cloud devices, tracing infrastructure, developer time, CI, device maintenance, and debugging. Repository license also does not automatically cover every model, dependency, cloud service, dataset, or app you connect.

If model selection is the part you are still deciding, Best AI Agent Models for Android: Tool APIs and GUI Specialists explains how to separate general tool APIs from GUI-focused models before you wire them into a phone runtime.

When Open-AutoGLM Fits

Open-AutoGLM is the strongest fit here when your project starts from visual phone-agent behavior. The Open-AutoGLM repository documents Android control through ADB, setup steps such as developer or USB debugging and ADB Keyboard, and model inference through either hosted model services or self-hosted serving. A local GPU is not mandatory if you use the hosted API path, but the model calls still have provider and availability conditions.

Use Open-AutoGLM when you want to study how a visual agent observes the phone, reasons about next steps, and executes through a device bridge. Its sensitive-operation confirmation and human takeover points for login or captcha-style situations are especially relevant for phone-agent safety work.

The practical cost is integration ownership. You need the device bridge, model endpoint, runtime environment, app-state handling, and recovery path to keep working. iOS is not a simple Android-equivalent switch; it uses its own WebDriverAgent setup. Choose Open-AutoGLM for visual agent research and framework-level control, not because open source magically removes deployment work.

When to Choose Mobilerun Framework or Cloud

Mobilerun is best when you need a developer-facing automation framework with inspectable execution. The Mobilerun repository, formerly known from the DroidRun lineage, documents CLI and Python control, accessibility trees, screenshots, model-provider selection, structured results, and saved trajectories. Android setup uses ADB, developer or USB debugging, and the Portal accessibility service.

Keep the framework and cloud product distinct. Mobilerun Framework runs the agent from your machine and lets you manage local phones, model providers, traces, and automation code. Mobilerun Cloud is a managed route with connected local phones plus hosted virtual or physical phones and API workflows. That can reduce infrastructure work, but it changes the operational dependency and cost model.

Mobilerun is a good choice if tracing matters. Its documented support for Arize Phoenix and Langfuse helps you inspect what the agent saw, what it decided, and where a run failed. That is valuable for teams building repeatable evaluations. For iOS, follow the project-specific Portal setup instead of assuming the Android setup transfers directly.

When Minitap mobile-use Fits

Minitap mobile-use fits teams that want natural-language mobile UI automation with structured extraction and configurable LLM providers. The mobile-use repository documents Android physical devices and emulators through ADB, an Android-only Docker quickstart, and structured mobile task patterns. It is not a consumer assistant; it is a framework path for builders who can manage devices, model providers, and runtime behavior.

The iOS boundary matters. The README’s manual device section lists iOS simulators on macOS and explicitly says physical iOS devices are not yet supported. Treat broader iOS wording as something to verify against the setup path you actually need. Simulator support, physical-device support, and Android support are different engineering commitments.

mobile-use is most attractive when your task can be expressed as a structured UI workflow and the target app exposes enough information for the framework to inspect. The repository also notes limits with games that lack accessibility-tree information. That is the kind of detail to check early, before building a workflow around an app surface the agent cannot reliably read.

Test One Reversible Task

Evaluate with one reversible task before adopting any framework. Do not begin with purchasing, sending, deleting, posting, or account-changing actions. Use a task such as opening a benign app screen, reading visible state, filling a draft field without submitting, or creating a disposable note that you can delete manually.

1. Define the phone state. Record device type, OS version, app version, connection method, permissions, and whether the app exposes accessibility information.
2. Fix the model route. Use one provider or self-hosted endpoint at a time, and note any token, hosting, or rate-limit cost.
3. Run the task once. Capture the prompt, screenshot or accessibility input, model response, action sequence, and final phone state.
4. Trigger one failure. Deny a permission, move to another screen, or remove a required field to see whether the framework stops, retries, or makes an unsafe assumption.
5. Inspect the trace. Check whether the failure was model reasoning, phone execution, app state, missing UI information, or infrastructure.

This is proposed evaluation, not a claim that any framework has passed a private test. For a deeper method, Android Phone Agent Benchmark Guide: Reliability, Safety, and Task Success gives a fuller way to evaluate reliability, recovery, and final state.

Choose a Ready-to-Use Android Path Instead

If you want phone help rather than framework maintenance, choose a product route instead of an open-source framework. FoneClaw is not an open-source phone-agent framework. It is our Android app with a configurable model path, a free default model, and governed supported tools for practical phone tasks such as opening apps, reading permitted SMS context, creating user-visible to-dos, and running approved supported actions.

That difference matters. With a framework, you own the repository setup, device bridge, model provider, tracing, upgrades, and recovery behavior. With FoneClaw, we focus on the ready-to-use Android runtime, clearer model selection, visible task progress, permissions, approvals, and recoverable phone workflows. Local phone execution still does not mean every model request stays on device, and not every Android action is an unattended automation. Scheduled unattended work is intentionally bounded.

Use FoneClaw when you want supported Android phone control without building the agent infrastructure yourself. The AI Agent Phone Control on Android: Intent, Confirmation, Action guide explains how intent, confirmation, and visible results fit together, while the FoneClaw Features page shows the current product surface. When you are ready to try the app route, the FoneClaw Download page lists current installation options.