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

Choose a Framework by the Job

The best open source phone agent frameworks solve different engineering problems. 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, then check the device bridge, model path, execution surface, license and maintenance load.

This is a selection guide, not a benchmark or report of hands-on testing. Mobilerun's October changes make device targeting and migration checks particularly relevant, but they do not establish that every framework has changed or that any project succeeds universally. Choose the stack whose operation and failure handling you can maintain.

Compare Setup, Execution, Models, and Licensing

Separate four layers before choosing. The framework is the code you run. Model inference is where reasoning happens. Phone execution is how taps, text input, screenshots or accessibility data reach the device. Inspection is how you determine what happened when a run stops or produces the wrong result.

Verify the license file for the revision you adopt, alongside model, dependency and service terms. A repository license does not automatically cover everything you connect. Open source also does not remove model API charges, hosted inference, cloud devices, tracing infrastructure, CI, device maintenance or debugging time.

If model selection is still unresolved, Best AI Agent Models for Android: Tool APIs and GUI Specialists explains how to distinguish general tool APIs from GUI-focused models before wiring them into a phone runtime.

When Open-AutoGLM Fits

Open-AutoGLM fits projects that begin with visual phone-agent behavior. The Open-AutoGLM repository documents Android control through ADB, setup such as developer or USB debugging and ADB Keyboard, and inference through hosted services or self-hosted serving. A local GPU is not mandatory for the hosted API route, but provider access and availability still apply.

Choose it when you want to study how a visual agent observes the phone, reasons about a next step and executes through a device bridge. Its sensitive-operation confirmation and human takeover points for login or captcha-style situations are relevant when evaluating how a workflow hands control back to a person.

The practical cost is integration ownership. You need the bridge, model endpoint, runtime environment, app-state handling and recovery path to keep working together. iOS is not an Android-equivalent switch: it requires its own WebDriverAgent setup. This makes Open-AutoGLM a framework-level choice for visual agent work, not a way to avoid deployment responsibilities.

When to Choose Mobilerun Framework or Cloud

Mobilerun fits developer-facing automation that needs inspectable execution. The Mobilerun repository, from the DroidRun lineage, documents CLI and Python control, accessibility trees, screenshots, provider selection, structured results and saved trajectories. Android uses ADB and Portal; follow the separate Portal setup for iOS.

Keep Framework and Cloud distinct. Framework runs the agent from your machine, where you manage devices, providers, traces and code. Cloud offers a managed route with connected local phones plus hosted virtual or physical phones and API workflows. It changes infrastructure responsibility and cost rather than making the local framework and hosted service interchangeable.

The October 14, 2026 release notes specify device-selection precedence for mobilerun run: explicit -d, then ANDROID_SERIAL, then the only online device. With multiple devices and no selection, it now errors instead of choosing the first. Inspect the connected-device list before running and deliberately identify the intended test phone.

The same release changes FastAgent execution: only the first function_calls block is used, later blocks are ignored, and a failed call skips the remainder of its block. Earlier actions may already have happened. Inspect the trace and actual phone state, fix the cause, then decide what remains to run. This is not automatic rollback or transactional execution.

For migration, the October 5 app-card and Portal changes require app_cards_dir/android/app_cards.json and app_cards_dir/ios/app_cards.json. A flat app_cards.json directly under app_cards_dir is ignored. Move cards into the correct platform folder and check Android identifiers separately from iOS bundle identifiers. Setup installs the appropriate pinned Portal release and leaves a newer installed Portal in place; do not substitute a guessed compatibility version.

Portal can restore its accessibility service and keyboard during an active run. If you intend to revoke that access, stop the agent first, then inspect the device's actual service and keyboard state. Do not disable security protections to keep a run going. Cloud UI-tree compatibility and the changed no-retry behavior for tap and type calls also belong in upgrade checks, not in assumptions of universal success.

For inspection, documented Arize Phoenix and Langfuse support can help you examine observations, decisions and failures. That is useful when building repeatable evaluations, provided you also verify the final device state.

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 a framework for builders who can manage devices, providers and runtime behavior, not a consumer assistant.

The iOS boundary matters. Its manual device setup lists iOS simulators on macOS and says physical iOS devices are not yet supported. Verify broader iOS wording against the setup path you need. Simulator support, physical-device support and Android support are different commitments.

mobile-use is most attractive when the task can be expressed as a structured UI workflow and the app exposes enough information to inspect. The repository notes limits with games that lack accessibility-tree information. Check that early rather than building around a surface the agent cannot adequately read. Provider flexibility does not compensate for missing device observations.

Test One Reversible Task

Use one harmless task to compare setup and inspection before adopting a framework. The following is a proposed, documentation-based check, not a claim that these projects passed a private test: on an owned Android test phone, open a chosen benign app and stop at its initial screen. Do not send messages, make payments, post content or change an account.

1. Identify the device. Inspect the connected-device list and record the intended phone's identity, OS, app version and connection. For Mobilerun, check explicit -d selection and any existing ANDROID_SERIAL value before running.
2. Confirm prerequisites. Check the framework's device bridge, actual permission and service state, provider endpoint and relevant app configuration. For Mobilerun, verify platform app-card paths after an upgrade.
3. Define a visible finish. Ask the agent to open the chosen app and stop. Success means the intended app is visibly open on the intended phone, not merely that the model reports completion.
4. Run once and inspect. Preserve the prompt, observations, model response, executed actions and final screen using the selected framework's available inspection tools.
5. Stop on a mismatch. Wrong-device selection, a failed action or an unexpected screen ends the check. Determine what already occurred before repairing the cause or retrying.
6. Check interruption separately. In a controlled follow-up, use a harmless changed starting screen to inspect stop and recovery behavior. If revoking Portal access, stop the active agent first and confirm the resulting device state.

Record whether a failure came from reasoning, execution, app state, missing UI information or infrastructure. Do not turn a successful app opening into a reliability or safety benchmark. Android Phone Agent Benchmark Guide: Reliability, Safety, and Task Success provides a fuller evaluation method when you need broader coverage.

Choose a Ready-to-Use Android Path Instead

If you want phone help rather than framework maintenance, consider a product route. FoneClaw is our ready-to-use Android phone agent, not an open-source framework. It provides a configurable model path and supported phone tools without requiring you to assemble one of these framework stacks.

Our configured model interprets and plans the request; enabled tools execute supported actions under actual Android permissions and the applicable tool approval policy. That distinction matters: model understanding is not execution authority, and an installed app does not provide universal control. Approval behavior depends on the policy and action rather than every step always producing a prompt.

With a framework, you own repository setup, device bridges, provider configuration, tracing, upgrades and recovery. With FoneClaw, the choice is whether our supported Android workflows meet your task. Local phone execution does not mean every model request stays on-device, and this alternative does not imply a native integration with Open-AutoGLM, Mobilerun or mobile-use.

The AI Agent Phone Control on Android: Intent, Confirmation, Action guide explains the relationship between requests, permissions and results. Review the FoneClaw Features page for the product's supported capabilities. Choose a framework when you need to build and maintain the execution stack; choose the ready-to-use route when supported phone tasks are the objective.