MobiAgent Wants to Make Mobile Agents Actually Customizable

You've probably seen a dozen demos of AI agents tapping through phone screens, and just as many claims that they're ready for real use. But if you've ever tried to build one yourself, you know the gap between a flashy demo and something you can actually deploy is enormous. MobiAgent is a framework from the IPADS-SAI team that tries to close that gap by giving you the full stack: a model family, an acceleration framework, and a benchmark, all under one roof.

What It Does

MobiAgent is a systematic framework for building customizable mobile agents. Rather than being a single model or a single tool, it's three things bundled together. First, there's MobiMind, an agent model family. Second, there's AgentRR, described as an agent acceleration framework. Third, there's MobiFlow, an agent benchmark. Together these cover the model, the runtime, and the evaluation side of the mobile agent problem.

The project has been moving fast since its initial open-source release in August 2025. The team has added multi-task execution support, a user profile memory system you can toggle with a flag (--user_profile on), an experience memory module, and support for custom workflows with a raw-data cleaning pipeline. There's also a unified GUI agent runner that supports one-click configuration of multiple models including MobiAgent itself, UI-TARS, AutoGLM, Qwen-VL, and Gemini. More recently they've pushed toward on-device inference, with a phone runner and an on-device inference framework called MobiInfer that targets both HarmonyOS and Android.

The model side includes several releases: MobiMind-Mixed variants capable of both Decider and Grounder tasks, reasoning-focused models like MobiMind-Reasoning-4B, and quantized versions such as MobiMind-1.5-2B-W8A8 for deployment on constrained hardware. There's also a GUI-based mobile "claw" project called MobiClaw, and a proactive on-device agent system called ClawMate.

Why It's Cool

• It covers the whole pipeline, not just the model. Most mobile agent projects hand you a model and wish you luck. MobiAgent gives you the model family, an acceleration framework, and a benchmark. That's the difference between a research artifact and something you can build a product around.

• On-device inference is a first-class concern. The December 2025 announcement about pure on-device inference on smartphones is the kind of thing that actually matters for real apps. You don't want to round-trip every screen tap to a cloud endpoint. MobiInfer targeting both HarmonyOS and Android suggests the team is thinking about deployment reality, not just benchmark scores.

• The unified runner is a practical touch. Being able to swap between MobiAgent, UI-TARS, AutoGLM, Qwen-VL, and Gemini with a one-click config is genuinely useful if you're evaluating which model fits your use case. It saves you from writing glue code for each one.

• Memory is treated as a real subsystem. Experience memory, user profile memory via Mem0/GraphRAG, and a dedicated MobiMem paper suggest the team understands that mobile agents need to remember context across sessions, not just within a single task.

• Custom workflows and data cleaning. The July 2026 addition of custom workflow support plus a raw-data cleaning pipeline is the unglamorous work that makes a framework actually usable. You can bring your own task structure and clean your own data.

How to Try It

1. Head to the repository: https://github.com/ipads-sai/mobiagent
2. Check out the system architecture diagram in the README to understand how MobiMind, AgentRR, and MobiFlow fit together.
3. For on-device inference, look at phone_runner/README.md to get started with running agents directly on a smartphone.
4. If you want to try different models without rewriting your setup, the unified runner is documented at the unify-runner branch README.
5. For multi-task scenarios, see runner/mobiagent/multi_task/README.md.
6. To enable the user profile memory system, pass --user_profile on when running the agent.

You can also grab the mobile app from the v1.0.1 release page, and download models from the Hugging Face collection linked in the README.

Final Thoughts

MobiAgent is clearly a research-driven project that's been steadily productized over the past year. The breadth is impressive, but it also means there's a lot of surface area to learn. If you're a developer who just wants a drop-in mobile agent, this might feel like more framework than you need. But if you're building mobile agent products and you care about customization, on-device performance, or benchmarking against real tasks, the combination of MobiMind, AgentRR, and MobiFlow gives you a foundation that's hard to find elsewhere. The frequent release cadence and the on-device push suggest the team isn't slowing down.

Follow @githubprojects for more developer tools and open source projects.