MCP Servers · Tongyi-MAI

MAI-UI

MAI-UI is a family of GUI agent foundation models (2B to 235B parameters) trained to understand and interact with mobile and desktop user interfaces. It supports multi-app workflows, device-cloud collaboration, and tool integration via MCP, achieving state-of-the-art performance on benchmarks like AndroidWorld (76.7%) and ScreenSpot-Pro (67.9%).

Talk to DEV.co Discuss Implementation

Source: GitHub — github.com/Tongyi-MAI/MAI-UI

1.8k

GitHub stars

178

Forks

Jupyter Notebook

Primary language

Apache-2.0

License (OSI-approved)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

Field	Value
Repository	Tongyi-MAI/MAI-UI
Owner	Tongyi-MAI
Primary language	Jupyter Notebook
License	Apache-2.0 — OSI-approved
Stars	1.8k
Forks	178
Open issues	32
Latest release	Unknown
Last updated	2026-04-20
Source	https://github.com/Tongyi-MAI/MAI-UI

What MAI-UI is

MAI-UI comprises vision-language models fine-tuned for GUI grounding and navigation tasks using large-scale reinforcement learning (up to 512 parallel environments, 50-step context). It features dynamic device-cloud execution selection, MCP tool augmentation for external service calls, and user interaction prompts via `ask_user` actions.

Quickstart

Get the MAI-UI source

Clone the repository and explore it locally.

terminalbash

1git clone https://github.com/Tongyi-MAI/MAI-UI.git2cd MAI-UI3# follow the project's README for install & configuration

Need it deployed, integrated, or customized instead? DEV.co ships production installs.

Best use cases

Mobile App Automation & RPA

Automate complex multi-app workflows on Android/iOS (e.g., shopping, booking, payment flows) with 2B–8B models on-device for low-latency, privacy-preserving execution.

Cross-App Task Orchestration

Execute tasks spanning multiple applications (calendar lookup, messaging, e-commerce) with intelligent app switching and MCP tool invocation for APIs (maps, transit, reservation systems).

Scalable Device-Cloud GUI Inference

Deploy smaller models on-device (2B/8B) for simple tasks; fall back to cloud models (32B/235B) for complex reasoning while minimizing API calls (~40% reduction observed) and protecting data sensitivity.

Implementation considerations

Requires vLLM (v0.11.0+) and transformers (v4.57.0+) for model serving; verify GPU/accelerator availability and memory (235B model requires significant VRAM).
MCP tool integration demands defining tool schemas and API endpoints; map existing backend services and authentication requirements before deployment.
Device-cloud routing logic must define task complexity thresholds and data sensitivity policies; unclear from docs how these are configured or tuned per use case.
RL environment scaling (512 parallel environments) implies substantial compute infrastructure; cost and latency trade-offs require profiling in staging.
Model weights (2B, 8B) are available; 32B and 235B availability and licensing terms for commercial use require clarification.

When to avoid it — and what to weigh

Requirement for Real-Time, Ultra-Low Latency — Model inference and device-cloud routing add latency; not suitable for subsecond reaction requirements or time-critical manual UI interactions.
Proprietary/Closed UI Systems Without Integration — Works best with standard Android/web UIs; integrating with proprietary or highly customized UI frameworks requires significant engineering effort.
No Need for Explainability or Audit Trails — If compliance/audit logging of agent decisions and reasoning is not required, the additional instrumentation overhead may not justify deployment.
Small Team or Limited ML/Infra Expertise — Deployment requires vLLM setup, device-cloud infrastructure, MCP configuration, and RL environment scaling; non-trivial for teams without ML Ops experience.

License & commercial use

Apache License 2.0 (Apache-2.0). Permissive OSI-approved license allowing commercial use, modification, and distribution with attribution and liability disclaimers.

Apache-2.0 permits commercial deployment. However, clarify: (1) whether larger model weights (32B, 235B) are similarly licensed; (2) any terms attached to Hugging Face or ModelScope hosting; (3) third-party dependencies' licenses (vLLM, transformers, etc.). Conduct full license audit before production.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

Signal	Assessment
Maintenance	Active
Documentation	Adequate
License clarity	Clear
Deployment complexity	High
DEV.co fit	Good
Assessment confidence	Medium

Security considerations

No explicit security audit or hardening details provided. Consider: (1) model inference may expose input UI state to cloud endpoints—review data residency/encryption; (2) MCP tool calls and external APIs introduce third-party risk; (3) device-cloud routing logic should validate task complexity to prevent unintended cloud fallback; (4) no mention of input sanitization or prompt injection mitigations; (5) accessibility API usage on mobile may expose sensitive UI elements. Recommend security review before handling sensitive workflows (banking, healthcare).

Alternatives to consider

Gemini-3-Pro / Claude Opus (API-based)

Closed-source, fully managed models with strong GUI reasoning; no device deployment, higher latency/cost, no device-cloud optimization, but simpler integration.

UI-Tars, Seed-1.8 (Other OSS GUI Agents)

Open-source competitors; MAI-UI claims SOTA on benchmarks, but trade-offs in model size flexibility, MCP integration, and device-cloud collaboration differ.

RPA / UiPath / Blue Prism

Mature, enterprise-grade RPA tools with drag-drop UI automation; no ML reasoning, higher setup cost, but proven stability and compliance support.

Software development agency

Build on MAI-UI with DEV.co software developers

Start with the 2B or 8B model on Hugging Face. Review the arXiv paper for architecture details. Contact the team ([email protected]) for enterprise deployment, licensing, or custom MCP integration.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

mcp-ui tuui shadcn-ui-mcp-server

Related on DEV.co

Explore the category and the services that help you build with it.

MCP Servers AI Application Development Mobile App Development Custom Software / Infrastructure Talk to DEV.co

MAI-UI FAQ

Can I run MAI-UI on mobile devices without a cloud connection?

Yes, 2B and 8B models are optimized for on-device deployment. 32B and 235B require cloud. Device-cloud framework detects task complexity and routes accordingly; offline operation depends on MCP tool availability.

How do I integrate custom tools or APIs with MCP?

MCP (Model Context Protocol) requires defining tool schemas. README links to websites but lacks detailed tool definition guide. Refer to arXiv paper or contact maintainers ([email protected]) for integration patterns.

What are the computational requirements?

2B/8B: Standard GPU (8–24GB VRAM). 32B: High-end GPU (40GB+). 235B: Multi-GPU or TPU cluster. vLLM enables batching and optimizations; exact throughput depends on context length and hardware.

Is MAI-UI suitable for production use?

Project is recent (created Dec 2025) with no formal release/SLA. Benchmark results are strong, but real-world stability, error recovery, and support SLAs are Unknown. Pilot in staging before production.

Software development & web development with DEV.co

Need help beyond evaluating MAI-UI? DEV.co is a software development agency offering software development services and web development for teams of every size. Our software developers and web developers build custom software, web applications, APIs, and mcp servers integrations — and maintain them long-term.

Evaluate MAI-UI for Your Automation Needs

Start with the 2B or 8B model on Hugging Face. Review the arXiv paper for architecture details. Contact the team ([email protected]) for enterprise deployment, licensing, or custom MCP integration.

Talk to DEV.co Schedule a Consultation