Vision-Only Desktop Agents vs. API-Based: Running the Same Task Both Ways

Mininglamp-AI released Mano-P, an open-source GUI-VLA agent project for edge devices that runs inference locally on Apple Silicon Mac mini and MacBook hardware with an M4 chip and 32GB or more RAM, or via a USB 4.0 compute stick. The release is phased: Mano-CUA Skills first for agent users such as OpenClaw and Claude Code, then local-side Mano-CUA models and SDK components, then the training, pruning and quantization methods behind the Mano-P models. Mininglamp-AI states that all CUA operations execute entirely on the local Mac mini and are not uploaded to external servers, with edge inference using INT8 activation quantization through its Cider SDK.

中文 https://github.com/Mininglamp-AI/Mano-P/blob/main/README CN.md | English 📖 Overview -overview | 🎯 Key Highlights -key-highlights | 🎬 Use Cases -use-case-demonstrations | 📊 Benchmark -benchmark-performance | 🔧 Skills -skills | 🤖 Models -models | ⚡ Inference SDK -inference-sdk | ⚗️ Approach -approach | 🌟 Technical Advantages -technical-advantages | 🔗 Applications -applications | 📄 Citation -technical-papers--citation | ❓ FAQ -faq Mano-P : "Mano" means "hand" in Spanish, and "P" stands for Private. We believe that both individuals and organizations can create their own Private AI, and a bright future of human-machine collaboration is on the horizon. Mano-P is a GUI-VLA agent project designed specifically for edge devices. It serves both as an open-source project and a hardware product solution. As an open-source project, Mano-P is being released in a phased, progressive manner, targeting three distinct groups of developers. In the first phase, we will open-source the Mano-CUA Skills. This phase is aimed at Agent enthusiasts—such as users of OpenClaw or Claude Code—enabling them to leverage the capabilities of Mano-CUA Skills to construct more intelligent CUA task workflows and overcome the bottlenecks associated with human intervention. In the second phase, we will open-source the local-side models and SDK components of Mano-CUA. This phase targets developers with high security requirements, allowing them to directly utilize GUI-VLA models capable of running inference locally on a Mac mini to build their own custom Skills, Tools, and more; crucially, all your CUA operations will be executed entirely on your local Mac mini and will not be uploaded to external servers. In the third phase, we will open-source the training methodologies and the pruning and quantization techniques used for the Mano-P models. This phase is designed for developers with specific model training needs, empowering them to apply our training methods to create their own on-device GUI-VLA models tailored to their unique requirements. Regarding our GUI-VLA models—which are capable of running inference directly on Mac mini and MacBook devices—we currently support two deployment methods: First, direct deployment on Mac mini or MacBook models equipped with an M4 chip and 32GB or more of RAM; and second, deployment utilizing a compute stick connected via a USB 4.0 port or higher. We will be releasing detailed instructions for both deployment methods in the near future, and we plan to expand our support to include additional deployment options in the future. - Complex GUI Automation : Autonomously complete complex interface operations containing hundreds of interactive elements - Cross-System Data Integration : Extract and integrate multi-source data through pure visual interaction without API interfaces - Long-Task Planning Execution : Support enterprise-level business process automation of dozens to hundreds of steps - Intelligent Report Generation : Automatically generate structured documents such as data analysis reports and work summaries - Edge-Native Inference : Efficient on-device execution on Apple Silicon via INT8 activation quantization Cider -inference-sdk - Autonomous Application Construction : Drives end-to-end software construction pipelines through visual GUI operation Mano-AFK -applications Mano-P builds upon the complete technical framework of the Mano project see Mano Technical Report https://arxiv.org/abs/2509.17336 , employing the Mano-Action bidirectional self-reinforcement learning method, three-stage progressive training SFT → Offline Reinforcement Learning → Online Reinforcement Learning , "think-act-verify" loop reasoning mechanism, and a closed-loop data circulation system to achieve high-precision GUI understanding and operation capabilities. The edge version is optimized through mixed-precision quantization, visual token pruning, and edge inference adaptation, enabling large-scale parameter models to run efficiently on edge devices like Mac mini/MacBook/computing sticks. - 1 on OSWorld Benchmark : Mano-CUA 1.1 achieves 58.2% success rate on OSWorld , ranking first among all specialized GUI agent models, outperforming the second-place opencua-72b 45.0% by 13.2 percentage points - Leading on WebRetriever Protocol I : Mano-CUA 1.1 scores 41.7 NavEval , surpassing Gemini 2.5 Pro Computer Use 40.9 and Claude 4.5 Computer Use 31.3 - Fully Local Execution : Runs inference locally on Apple M4 chip with 32GB RAM Mac mini or MacBook . No cloud API calls required. All screenshots and task data stay on-device - High-Performance Inference : Mano-CUA-4B achieves ~80 tokens/s decode on Apple M5 Pro; with Cider's W8A8 activation quantization, prefill speeds up by ~12.7% over the W8A16 baseline - Autonomous Long-Task Execution : Supports complex business processes with end-to-end automation without internet connectivity - Edge-Native INT8 Acceleration : Companion Cider -inference-sdk SDK adds the W8A8 / W4A8 activation-quantization primitives MLX lacks natively, delivering 1.4x–2.2x prefill speedup over MLX W4A16 on Apple M5 Pro — works with any MLX model, not just Mano-P - Autonomous Software Construction : Mano-AFK -applications drives a full PRD → code → deploy → test → fix loop using Mano-P as its local vision model for real-browser E2E testing — from a single natural-language prompt to a deployed, tested application, no human in the loop Mano-P AFK EN.mp4 We demonstrated the fully automated application construction process of Mano-AFK -applications . After receiving natural language requirements, the system sequentially completes requirement clarification, technical architecture design, code generation, local deployment, and multi-level testing API interface testing, LLM based page visual inspection, and end-to-end GUI automation testing driven by VLA model . When the test fails, the system automatically locates the root cause of the problem, fixes the code, and deploys verification again, iterating until all test cases pass. The entire process does not require manual intervention, and ultimately delivers a runnable application with complete requirement documents and build reports. Mano-P.Commercial video intelligent system EN.mp4 We fully demonstrated the actual workflow of a commercial video intelligent system. Starting from the user's command, the system automatically completes the entire process of video generation, uploading, analysis, editing, and secondary evaluation. During the process, the system can autonomously operate web pages and editing software, complete fine operations such as file processing and subtitle modification, and generate analysis reports containing subjective evaluations and objective indicators. By comparing the differences between the initial and refined versions, visually present the overall capabilities and application effects of the system. Mano-P. EN.mp4 Mano-P, The small-sized end side GUI-VLA model can run directly on your computer, supporting direct inference operation on Macmini/Macbook with M4 chip and above, as well as direct operation on plug and play computing power sticks. In the CUA scenario, break through the bottleneck of human participation in the Agent workflow. Mano-P, The first step in leading Private AI. Mano . EN.mp4 Mano-P excels not only in enterprise-level business automation but also integrates seamlessly into daily life. This video demonstrates the system's application in Mahjong gameplay: through pure visual understanding of the game interface, it autonomously completes tile recognition, analysis, and decision-making. This case validates Mano-P's general-purpose capabilities beyond work scenarios—from office automation to leisure entertainment, from structured data processing to unstructured game interactions, truly realizing the vision of "Private AI." One model, adapting to every aspect of life and work. 0429 Mano-AFK+Cider EN compressed.mp4 This video demonstrates the combined capability of Mano-AFK -applications and the Cider -inference-sdk inference acceleration SDK. Starting from a single natural-language requirement, Mano-AFK autonomously performs requirement clarification, architecture design, code generation, and local deployment. In the E2E testing stage it drives a real browser through the local Mano-P vision model — accelerated by Cider — to run GUI automation tests; on failure, it localizes the defect, patches the code, and re-verifies until a runnable application is delivered. Cider supplies the INT8 activation-quantization primitives that give Mano-P a substantial prefill speedup on Apple Silicon, so the entire build–test–fix loop runs fully on-device, balancing autonomy, privacy, and performance. Performance of the Mano series models in multiple benchmarks: 📊 Expand Evaluation Data | Models | Protocol | CA | | CV | | PAR | | Saliency | | | | | |---|---|---|---|---|---|---|---|---|---|---|---|---| | | | Acc | F1 | Acc | F1 | Acc | F1 | KL↓ | CC↑ | SIM↑ | NSS↑ | AUC↑ | | Random | P1 | 10.42 | 11.03 | 10.76 | 10.95 | 15.94 | 16.00 | 2.1789 | 0.0452 | 0.2852 | 0.1081 | 0.5340 | | | P2 | 10.01 | 10.74 | 10.32 | 10.50 | 14.39 | 15.04 | 4.3378 | 0.0270 | 0.2274 | 0.0665 | 0.5273 | | Zero-shot for MLLMs | | | | | | | | | | | | | | GPT4o | P1 | 15.17 | 6.57 | 16.11 | 9.58 | 16.71 | 10.34 | 1.9423 | 0.4660 | 0.4602 | 1.2842 | 0.7848 | | | P2 | 10.26 | 4.77 | 12.16 | 7.66 | 15.00 | 8.55 | 2.2650 | 0.4097 | 0.4028 | 1.2418 | 0.7807 | | Gemini 2.0 Flash | P1 | 17.18 | 5.13 | 25.06 | 8.39 | 24.94 | 9.52 | 1.4726 | 0.3380 | 0.3751 | 0.8629 | 0.7296 | | | P2 | 10.45 | 4.26 | 12.60 | 4.95 | 15.96 | 7.90 | 1.6373 | 0.3542 | 0.3490 | 1.0027 | 0.7590 | | GPT-5.2 | P1 | 17.83 | 7.67 | 22.22 | 12.55 | 16.17 | 9.74 | 1.3262 | 0.4852 | 0.4632 | 1.3078 | 0.7969 | | | P2 | 15.31 | 5.14 | 19.88 | 10.27 | 13.56 | 7.42 | 1.5444 | 0.4379 | 0.4092 | 1.3006 | 0.7999 | | Claude Sonnet 4.5 | P1 | 10.34 | 5.8 | 13.26 | 9.84 | 16.02 | 9.94 | 1.4235 | 0.4912 | 0.4213 | 1.2956 | 0.8042 | | | P2 | 10.34 | 5.55 | 13.27 | 7.08 | 16.02 | 9.6 | 1.2855 | 0.4564 | 0.4781 | 1.3112 | 0.7915 | | Llama 4 Scout | P1 | 13.98 | 9.96 | 10.25 | 6.51 | 13.27 | 8.11 | 3.7166 | 0.3331 | 0.3849 | 0.8828 | 0.7238 | | | P2 | 10.00 | 7.33 | 11.10 | 8.49 | 14.35 | 7.42 | 3.7434 | 0.3019 | 0.3452 | 0.8848 | 0.7258 | | Qwen2.5-VL-7B | P1 | 15.88 | 5.21 | 10.07 | 6.07 | 12.26 | 4.96 | 12.0586 | 0.0999 | 0.2154 | 0.2578 | 0.5852 | | | P2 | 10.25 | 3.95 | 10.89 | 5.83 | 14.39 | 5.73 | 12.7596 | 0.0762 | 0.1855 | 0.2195 | 0.5753 | | InternVL3-8B | P1 | 13.35 | 7.78 | 14.71 | 8.02 | 10.20 | 6.95 | 12.6480 | 0.0572 | 0.1895 | 0.1140 | 0.5769 | | | P2 | 10.58 | 6.70 | 10.94 | 8.12 | 12.68 | 6.32 | 12.1385 | 0.0604 | 0.1819 | 0.1395 | 0.5859 | | Fine-tune for MLLMs | | | | | | | | | | | | | | Qwen2.5-VL-7B | P1 | 22.51 | 19.11 | 23.39 | 10.83 | 32.06 | 25.88 | 1.5091 | 0.6953 | 0.6118 | 1.8937 | 0.8579 | | | P2 | 13.72 | 13.25 | 13.03 | 10.94 | 21.24 | 20.65 | 2.2496 | 0.5359 | 0.4793 | 1.6439 | 0.8221 | | InternVL3-8B | P1 | 20.94 | 18.41 | 21.96 | 11.02 | 30.33 | 24.66 | 1.2551 | 0.7014 | 0.6340 | 1.9896 | 0.8670 | | | P2 | 12.81 | 11.83 | 12.16 | 11.11 | 19.26 | 19.27 | 1.8759 | 0.6282 | 0.5467 | 2.0621 | 0.8627 | | Mano-CUA 1.1 | | | | | | | | | | | | | | Stage I | P1 | 31.27 | 30.53 | 27.31 | 25.18 | 35.16 | 34.45 | 0.6794 | 0.7670 | 0.7015 | 2.1347 | 0.8710 | | | P2 | 21.89 | 22.06 | 18.27 | 18.57 | 23.77 | 23.87 | 1.5759 | 0.6482 | 0.6167 | 2.1021 | 0.8627 | | Stage II | P1 | 32.59 | 31.46 | 27.57 | 25.76 | 37.73 | 35.79 | 0.6736 | 0.7686 | 0.7120 | 2.1688 | 0.8853 | | | P2 | 20.55 | 21.26 | 15.37 | 15.15 | 25.36 | 25.83 | 0.5617 | 0.6440 | 0.6130 | 2.1090 | 0.8602 | | Stage III | P1 | 34.58 | 33.99 | 31.92 | 28.37 | 39.42 | 37.63 | 0.6073 | 0.7853 | 0.7248 | 2.2103 | 0.8938 | | | P2 | 25.29 | 25.83 | 20.21 | 19.29 | 26.49 | 26.54 | 1.4617 | 0.6725 | 0.6330 | 2.1788 | 0.8776 | | Dataset | Method | Saliency | | | | | |---|---|---|---|---|---|---| | | | KL↓ | CC↑ | SIM↑ | NSS↑ | AUC↑ | | MIT1003 | FastSal | 1.036 | 0.590 | 0.478 | 2.008 | 0.875 | | | SAM-Resnet | 1.247 | 0.746 | 0.597 | 2.752 | 0.902 | | | DAV | 0.753 | 0.699 | 0.566 | 2.574 | 0.897 | | | UNISAL | 1.014 | 0.734 | 0.597 | 2.759 | 0.902 | | | Transalnet | 0.660 | 0.722 | 0.592 | 2.631 | 0.903 | | | SUM | 0.563 | 0.768 | 0.630 | 2.839 | 0.913 | | | Mano-CUA 1.1 | 0.648 | 0.770 | 0.698 | 2.950 | 0.902 | | SalECI | SSM | 0.720 | 0.599 | 0.611 | 1.396 | 0.830 | | | DeepGaze IIE | 0.995 | 0.560 | 0.399 | 1.327 | 0.842 | | | EML-NET | 1.220 | 0.510 | 0.536 | 1.232 | 0.807 | | | Transalnet | 0.873 | 0.717 | 0.534 | 1.723 | 0.824 | | | Temp-Sal | 0.712 | 0.719 | 0.629 | 1.768 | 0.813 | | | SSwinTransformer | 0.652 | 0.687 | 0.606 | 1.701 | 0.868 | | | Mano-CUA 1.1 | 0.615 | 0.769 | 0.695 | 1.735 | 0.868 | | Methods | Saliency | | | | |---|---|---|---|---| | | CC ↑ | SIM ↑ | NSS ↑ | AUC ↑ | | ACLNet | 0.477 | 0.329 | 2.36 | 0.915 | | TASED-Net | 0.479 | 0.366 | 2.63 | 0.916 | | STAViS | 0.569 | 0.425 | 2.94 | 0.931 | | ViNet | 0.569 | 0.409 | 3.06 | 0.928 | | CASP-Net | 0.620 | 0.478 | 3.34 | 0.940 | | Mano-CUA 1.1 | 0.642 | 0.481 | 2.99 | 0.929 | | | Emotion Valence | | Emotion Arousal | | |---|---|---|---|---| | | Acc ↑ | Acc ± 1 ↑ | Acc ↑ | Acc ± 1 ↑ | | Qwen2.5-VL-7B | 13.3 | 38.1 | 10.8 | 35.5 | | Mano-CUA 1.1 | 20.2 | 46.5 | 18.7 | 47.3 | 📊 Expand Evaluation Data Comparison of Task Execution Success Rate SR on Online-Mind2Web Benchmark Avg. Tokens/img represents the average visual token retention rate per image; lower values indicate more aggressive pruning. GSPruning is a novel token pruning method designed for Vision-Language Models to efficiently process high-resolution web interfaces by preserving global spatial structure through anchor points and identifying semantic outliers for critical UI elements. It achieves 2-3× throughput speedup with minimal performance loss, enabling more efficient autonomous web agents. | Model | Method | Avg. Tokens /img ↓ | Training samples/s ↑ | SR ↑ | |---|---|---|---|---| | Qwen3VL-2B | Baseline w/o FT | 100% | 5.08 | 0.290 | | | Baseline FT | 100% | 5.09 | 0.390 | | | TextGuide | 12.55% | 13.54 | 0.310 | | | FlashVLM 4 | 12.55% | 17.01 | 0.343 | | | Compressor-VLA 11 | 13.33% | 16.92 | 0.293 | | | HiPrune 16 | 25.09% | 16.67 | 0.333 | | | PDrop 33 | 41.47% | 10.43 | 0.330 | | | IVC | 25.09% | 7.89 | 0.303 | | | Mano-CUA 1.1 | 25.09% | 20.04 | 0.370 | | | Mano-CUA 1.1 | 12.57% | 22.62 | 0.336 | | Qwen3VL-4B | Baseline FT | 100% | 3.24 | 0.425 | | | PDrop | 41.47% | 5.58 | 0.365 | | | IVC | 25.09% | 4.67 | 0.343 | | | GSPruning | 25.09% | 16.72 | 0.400 | 📊 Expand Evaluation Data Comparison of four inference configurations on 100 real-machine macOS GUI tasks MacBook Pro · Apple M5 · 16GB : Overall Metrics | Metric | Mano-CUA Cloud | Mano-CUA-1.0-4B | Mano-CUA-2.0-4B | Qwen3-VL-Plus | |---|---|---|---|---| | Pass rate | 83.0% | 47.0% | 56.0% | 39.0% | | Avg time per step | 9.3s | 8.0s | 7.9s | 10.2s | Difficulty Tiers | Tier | Tasks | Mano-CUA Cloud | Mano-CUA-1.0-4B | Mano-CUA-2.0-4B | Qwen3-VL-Plus | |---|---|---|---|---|---| | A | 25 | 23/25 92% | 21/25 84% | 22/25 88% | 18/25 72% | | B | 45 | 37/45 82% | 18/45 40% | 24/45 53% | 14/45 31% | | C | 30 | 23/30 77% | 7/30 23% | 10/30 33% | 7/30 23% | Per Category | Category | Mano-CUA Cloud | Mano-CUA-1.0-4B | Mano-CUA-2.0-4B | Qwen3-VL-Plus | |---|---|---|---|---| | Browser / Web 31 | 28 90% | 23 74% | 21 68% | 18 58% | | Fuzzy descriptions 10 | 8 80% | 3 30% | 3 30% | 3 30% | | File management 7 | 5 71% | 3 43% | 3 43% | 4 57% | | WeChat 6 | 5 83% | 2 33% | 4 67% | 2 33% | | WeCom / Feishu / DingTalk 6 | 6 100% | 2 33% | 5 83% | 3 50% | | System settings 6 | 3 50% | 3 50% | 5 83% | 3 50% | | WPS / Office 5 | 5 100% | 0 0% | 2 40% | 0 0% | | No open hint 5 | 4 80% | 1 20% | 3 60% | 2 40% | | Notes / Reminders 4 | 4 100% | 2 50% | 2 50% | 0 0% | | System utilities 3 | 3 100% | 3 100% | 3 100% | 1 33% | | Long chains 10 | 8 80% | 3 30% | 3 30% | 2 20% | | Cross-app 5 | 3 60% | 0 0% | 1 20% | 0 0% | Mano-CUA 1.1 + Bash Tool With shell tools enabled in Mano-CUA 1.1, cloud mode can use system commands to handle tasks that pure GUI struggles with: | ID | Task | GUI Only | +Bash | Method | |---|---|---|---|---| | 55 | Finder: tag red label | Fail | Pass | GUI right-click + tag | | 57 | Volume to 50% | Fail | Pass | osascript 'set volume' | | 58 | Max mouse pointer size | Fail | Pass | defaults write | | 82 | Calculator → TextEdit | Fail | Pass | GUI Calculator + TextEdit | | 86 | Translate & save to txt | Fail | Pass | echo to file | | 99 | Rotate image | Fail | Pass | sips -r 90 | Mano-CUA 1.1+Bash pass rate: 90/100 = 90% +7; 2 of those were false negatives caused by a system proxy misconfiguration in the previous run . Why Local Matters - ✅ Faster per step : 7.9s vs Mano-CUA 1.1 9.3s vs Qwen 10.2s - ✅ Fully local : zero outbound traffic for screenshots or task descriptions, no network dependency - ✅ Lightweight deployment : ~6.4 GB memory footprint, runs on a MacBook - ✅ Small model beats large model : Mano-CUA-4B 56% significantly outperforms the cloud-based general-purpose VL model Qwen3-VL-Plus 39% as a fully local 4B model, demonstrating the value of GUI-specialized fine-tuning Current Gap The 4B on-device model reaches 56% vs Mano-CUA 1.1's 83%. The gap concentrates in fuzzy descriptions, cross-app workflows, and deep office-suite operations — these are the explicit directions for the next iteration. Qwen3-VL-Plus, as a cloud-based general-purpose VL model, only reaches 39%, primarily limited by Chinese input focus issues, poor adaptation to non-browser apps, and step-limit truncation — showing that general VL capability does not equal GUI Agent capability. Mano-CUA 1.1+Bash mode pushes the pass rate to 90% via osascript / sips / defaults write , demonstrating the ceiling of a hybrid GUI + Shell strategy. Test Configuration - Hardware: MacBook Pro · Apple M5 · 16GB RAM - Cloud model: Claude Sonnet 4.5 via mano.mininglamp.com - Local model: Mano-CUA-4B W8A16, MLX - Qwen model: Qwen3-VL-Plus via llm-gateway.mlamp.cn - Task set: 100 tasks covering browser/web, app operations, long chains, cross-app, fuzzy descriptions Mano-Skill is a desktop GUI automation tool based on the Mano model, driving cross-platform graphical interface operations through natural language. We provide two different usage forms for the same core capability to adapt to different usage scenarios and user groups. - Natural Language Driven : Users describe tasks in natural language, and the system automatically executes GUI operations - Flexible Inference Modes : - Local Mode : Models run locally, data stays on device, fast response - Run directly on Mac mini/MacBook M4 chip or above, 32GB+ RAM - Or use Mano-P computing stick via USB 4.0 connection - Cloud Mode : Without local model configuration, uses cloud API service mano.mininglamp.com - System automatically detects local model configuration and seamlessly switches inference modes - Comprehensive Interaction Support : Click, type, hotkey, scroll, drag, mouse movement, screenshot, wait, app launch, URL navigation - Cross-Platform Support : macOS stable , Windows, Linux Beta Cloud Mode Default 1. Capture current screen screenshot 2. Send screenshot and task description to cloud vision model mano.mininglamp.com 3. Cloud model analyzes and returns next action instruction 4. Local client executes operation click, type, etc. 5. Loop execution until task completion Local Mode 1. Capture current screen screenshot 2. Run Mano-P model on local device Mac mini/MacBook or computing stick for inference 3. Local model analyzes and returns next action instruction 4. Client executes operation click, type, etc. 5. Loop execution until task completion Cloud Mode: - ⚠️ Data Sent : Screenshots and task descriptions sent to mano.mininglamp.com for real-time visual analysis - ✅ Data Not Sent : Does not access or transmit local files, clipboard contents, system credentials - ⚠️ Privacy Note : Avoid displaying sensitive documents, chat logs, or credential information on screen when running tasks Local Mode Mac mini/MacBook or Computing Stick : - ✅ Fully Local Processing : All data processing is completed locally, screenshots and task descriptions never leave the device - ✅ Data Stays on Device : Does not access or transmit any data to external servers - ✅ Maximum Privacy Protection : Suitable for handling sensitive information and high-security scenarios General Assurance: - ✅ Open Source Auditable : Complete source code publicly available for review If you want to use Mano-P directly to accomplish GUI automation tasks, here are two different usage forms. Choose the one that best fits your use case. Use Case : For human users — developers and advanced users invoking mano-cua directly in a terminal to quickly execute one-off or scripted GUI automation tasks Installation : Install via Homebrew brew tap Mininglamp-AI/tap brew install mano-cua The installation process will automatically: - Create an isolated Python 3.13 virtual environment - Install required dependencies including Tkinter GUI library - Configure the executable command to system PATH Usage : Cloud mode default — no extra setup required mano-cua run "Open WeChat and tell FTY the meeting is postponed" mano-cua run "Search for AI news on Xiaohongshu and display the first post" Stop current task mano-cua stop mano-cua ships with two inference modes: cloud and local . Cloud is the default; to run Mano-P locally on macOS Apple Silicon, use the --local flag: First-time local setup: verify env / install SDK / pull the local model mano-cua check mano-cua install-sdk mano-cua install-model Run a task in local mode mano-cua run "Open Safari and search for Python" --local mano-cua run "Type hello in the search box" --local --url "https://www.baidu.com" --minimize --max-steps 15 In local mode, Mano-P runs on-device via MLX — screenshots and task descriptions make zero network calls for inference. Features : - ✅ Command-line interface, quick invocation - ✅ Virtual environment isolation, no system Python pollution - ✅ Suitable for script integration and batch processing - ✅ Can be embedded in shell scripts - ✅ Cloud / local inference modes — --local switches to on-device in one flag Installation & Distribution : - Homebrew Tap : github.com/Mininglamp-AI/homebrew-tap https://github.com/Mininglamp-AI/homebrew-tap Use Case : For AI agents — Claude Code, OpenClaw and similar agents autonomously invoke GUI automation capabilities mid-reasoning to complete user tasks, no manual command execution needed Installation : Option 1: Install via Claude Code In Claude Code, skills exist as "commands". Installation steps: 1. Download the skill zip package from ClawHub https://clawhub.ai/hanningwang/mano-cua 2. After extraction, copy files to Claude Code's commands directory 3. Restart Claude Code or in a new session, the skill will be automatically available Option 2: Install via ClawHub CLI Recommended Use the ClawHub CLI tool for one-click installation and skill management: Install skill clawhub install mano-cua Install specific version clawhub install mano-cua --version 1.0.0 Update skill to latest version clawhub update mano-cua After installation, start a new Claude Code or OpenClaw session to use. Prerequisites : ClawHub CLI tool must be installed first. See: OpenClaw Documentation - ClawHub https://docs.openclaw.ai/tools/clawhub Usage : When users make requests to AI agents that require GUI operations, the agent will automatically invoke this skill: User: "Help me open WeChat, find FTY's chat window, and tell him the meeting is postponed to tomorrow" Agent: Automatically invokes mano-skill to complete GUI operation Features : - ✅ Autonomously invoked by AI agents, no manual command execution needed - ✅ Deeply integrated with agent reasoning capabilities - ✅ Suitable for complex multi-step task automation - ✅ ClawHub ecosystem with version management and security scanning Project Resources : - Source Code : github.com/Mininglamp-AI/mano-skill https://github.com/Mininglamp-AI/mano-skill - ClawHub Home : clawhub.ai/hanningwang/mano-cua https://clawhub.ai/hanningwang/mano-cua - Version : v1.0.0 - License : MIT - Screen Recording Permission - Accessibility Permission keyboard/mouse control - Grant permissions in System Preferences → Privacy & Security - Sensitive or potentially dangerous operations require user confirmation before execution - Users can stop tasks at any time - Only one task can run on each device simultaneously - Only supports primary display multi-display environment When a task is running, a small status panel appears in the top-right corner of the screen to: - Display real-time task status and progress - Provide task management functions pause/stop - Remind users that an automation task is running to avoid accidental interference Beta Version Notice : Mano-Skill is currently in Beta testing phase. - macOS : ✅ Preferred and most thoroughly tested platform, stable and ready for use - Windows and Linux :⚠️ Platform adaptations not yet fully completed, minor issues may occur We are continuously improving cross-platform compatibility. Feedback is welcome. If you want to integrate Mano-P's model capabilities into your own applications, this section provides performance metrics and usage guidelines. The table below presents actual inference benchmark results of Mano-CUA-4B running on Apple M5 Pro with the Cider inference SDK. Using W8A16 MLX's native weight-only quantization path as the baseline — the same reference convention adopted in Cider's quantization benchmark -inference-sdk — enabling Cider's W8A8 activation quantization reduces prefill time from 2.839s to 2.519s on the same input, a ~12.7% prefill speedup . For more data, refer to the ⚡ Inference SDK -inference-sdk section below. | Model | Chip | Bandwidth | Framework | Context Length | Quantization | Prefill Time s | Decode Speed tokens/s | |---|---|---|---|---|---|---|---| | Mano-CUA-4B | Apple M5 Pro 64GB RAM | 307 GB/s | Cider | 4516 | W8A16 | 2.839 | 80.1 | | | | | | | W8A8 | 2.519 | 79.5 | Model Download: | Platform | Base Model | MLX 8-bit Quantized | |---|---|---| | 🤗 Hugging Face | Mano-CUA-2.0-4B https://huggingface.co/Mininglamp-2718/Mano-CUA-2.0-4B | Mano-CUA-2.0-4B-MLX-8bit https://huggingface.co/Mininglamp-2718/Mano-CUA-2.0-4B-MLX-8bit | | 🪄 ModelScope CN | Mano-CUA-2.0-4B https://www.modelscope.cn/models/Mininglamp2718/Mano-CUA-2.0-4B | Mano-CUA-2.0-4B-MLX-8bit https://www.modelscope.cn/models/Mininglamp2718/Mano-CUA-2.0-4B-MLX-8bit | | 🪄 ModelScope AI | Mano-CUA-2.0-4B https://www.modelscope.ai/models/Mininglamp2718/Mano-CUA-2.0-4B | Mano-CUA-2.0-4B-MLX-8bit https://www.modelscope.ai/models/Mininglamp2718/Mano-CUA-2.0-4B-MLX-8bit | Cider is an inference acceleration SDK developed on top of MLX for macOS. It provides online activation quantization operators absent in MLX, with custom int-matmul kernels built as MLX custom primitives supporting full lazy evaluation. It also includes service-side extensions and non-intrusive compatibility patches for mlx vlm validated on mlx vlm 0.4.3 , including fixes for Qwen3-VL multi-image inference issues related to RoPE position handling and chunked prefill. Cider uses conditional compilation : the INT8 TensorOps C++ extension is only built on Apple M5+. | Chip | pip install -e . behavior | import cider behavior | |---|---|---| | M5+ | Full build CMake + Metal kernels | All features available | | M4 and below | Skips C++ build, installs pure-Python package | is available → False, convert model is a warning no-op | Override via environment variable: CIDER FORCE BUILD=1 pip install -e . Force build e.g., CI CIDER FORCE BUILD=0 pip install -e . Force skip | Mode | Weights | Activations | Compute Path | Status | |---|---|---|---|---| | W8A8 | INT8 symmetric | INT8 per-token | TensorOps matmul2d | ✅ Implemented | | W4A8 | INT4 packed uint8 | INT8 per-token | Unpack → TensorOps | ✅ Implemented | | W4A16 | — | — | MLX built-in | Baseline | | W8A16 | — | — | MLX built-in | Baseline | W4A16 and W8A16 are already supported by MLX natively — this SDK provides the missing W8A8 and W4A8 modes that MLX does not implement. MLX's quantization is weight-only : QuantizedLinear dequantizes weights to FP16 and uses FP16 GEMM. While MLX's Steel NAX templates are generic enough to be instantiated with INT8 types and would achieve identical raw matmul throughput — see the transparent benchmark https://github.com/Mininglamp-AI/cider/blob/main/benchmarks/mlx native/cider vs mlx int8.md , MLX does not provide the quantization/dequantization pipeline needed for actual W8A8 inference. Cider fills this gap with fused quantize-matmul-dequant primitives, implementing online INT8 activation quantization and INT8 TensorOps-based compute for the supported inference paths. | Granularity | Description | Speed | Precision | |---|---|---|---| | Per-channel | One scale per output channel | Fastest 1.8x prefill | Slightly lower | | Per-group gs=128 | One scale per 128 elements | Fast 1.5x prefill | Moderate precision retention | | Per-group gs=64 | One scale per 64 elements | Moderate 1.3x prefill | Higher precision | Individual Operator Latency Shape N=10240, K=2560 | M | PC ms | PG ms | w8a16 | w4a16 | PC/w8 | PC/w4 | PG/w8 | PG/w4 | |---|---|---|---|---|---|---|---|---| | 1 | 0.27ms | 0.26ms | 0.26ms | 0.18ms | 0.96x | 0.67x | 0.99x | 0.69x | | 128 | 0.34ms | 0.39ms | 0.49ms | 0.44ms | 1.43x | 1.28x | 1.26x | 1.13x | | 1024 | 1.23ms | 1.52ms | 2.24ms | 2.04ms | 1.82x | 1.66x | 1.47x | 1.34x | | 4096 | 4.41ms | 5.65ms | 8.12ms | 7.72ms | 1.84x | 1.75x | 1.44x | 1.37x | | 8192 | 8.71ms | 11.40ms | 16.23ms | 15.09ms | 1.86x | 1.73x | 1.42x | 1.32x | Shape N=2560, K=10240 | M | PC ms | PG ms | w8a16 | w4a16 | PC/w8 | PC/w4 | PG/w8 | PG/w4 | |---|---|---|---|---|---|---|---|---| | 1 | 0.25ms | 0.26ms | 0.26ms | 0.20ms | 1.03x | 0.78x | 0.98x | 0.75x | | 128 | 0.39ms | 0.41ms | 0.55ms | 0.46ms | 1.43x | 1.19x | 1.35x | 1.12x | | 1024 | 1.31ms | 1.65ms | 2.35ms | 2.14ms | 1.80x | 1.64x | 1.43x | 1.30x | | 4096 | 5.37ms | 6.79ms | 8.54ms | 8.04ms | 1.59x | 1.50x | 1.26x | 1.18x | | 8192 | 10.97ms | 12.94ms | 17.28ms | 16.23ms | 1.58x | 1.48x | 1.34x | 1.25x | End-to-End VLM Qwen3-VL-2B | Prompt Tokens | FP16 Prefill tok/s | W8A16 Prefill tok/s | W8A8 PC Prefill tok/s | FP16 Decode tok/s | W8A16 Decode tok/s | W8A8 PC Decode tok/s | |---|---|---|---|---|---|---| | 1334 | 3010 | 2065 | 3242 | 70 | 107 | 104 | | 2393 | 2868 | 1847 | 2983 | 69 | 97 | 100 | | 3455 | 2777 | 1741 | 2796 | 66 | 90 | 95 | Qwen3-VL-4B | Prompt Tokens | FP16 Prefill tok/s | W8A16 Prefill tok/s | W8A8 PC Prefill tok/s | FP16 Decode tok/s | W8A16 Decode tok/s | W8A8 PC Decode tok/s | |---|---|---|---|---|---|---| | 1334 | 1884 | 1786 | 2186 | 32 | 56 | 54 | | 2393 | 1815 | 1700 | 2028 | 31 | 55 | 52 | | 3455 | 1755 | 1603 | 1881 | 30 | 52 | 49 | LLM Quantization: Precision vs. Speed Comparison | Models | Quantization Configuration | wikitext2 PPL ↓ | Prefill Time s ↓ | Peak Memory GB ↓ | |---|---|---|---|---| | Qwen3-8B | FP16 | 9.726 | 179.9 | 18.93 | | | W8A16 mlx RTN | 9.707 | 221.3 | 12.07 | | | W8A8 per-channel | 9.756 | 123.5 | 11.32 | | | W8A8 per-group gs=64 | 9.744 | 179.1 | 11.83 | | | W8A8 per-group gs=128 | 9.727 | 165.8 | 11.61 | | Llama3-8B | FP16 | 6.138 | 175.8 | 18.32 | | | W8A16 mlx RTN | 6.147 | 236.9 | 11.46 | | | W8A8 per-channel | 6.271 | 123.3 | 10.69 | | | W8A8 per-group, gs=64 | 6.269 | 178.7 | 11.19 | | | W8A8 per-group, gs=128 | 6.270 | 155.7 | 10.98 | - 🔗 Repository: github.com/Mininglamp-AI/cider https://github.com/Mininglamp-AI/cider If you are a researcher or wish to train customized GUI Agent models based on your own data, we plan to open-source the complete Mano-Action training methodology and related tools. Release Soon Mano-Action is a bidirectional self-reinforcement training framework specifically designed for GUI Grounding. Unlike traditional unidirectional prediction methods, Mano-Action achieves more robust interface understanding through Text↔Action cycle consistency learning, enabling the model to master both "locating elements from descriptions" and "describing given elements" simultaneously. - Bidirectional Cycle Learning : Mutual reinforcement between Text → Action and Action → Text - Three-Stage Progressive Training : Supervised Learning → Offline RL → Online RL - Closed-Loop Data Generation : Automatically generate high-quality training data for continuous model improvement - Edge Optimization Adaptation : Includes quantization, pruning, and other edge deployment optimization techniques - 🎓 Academic Research : Explore new approaches to GUI understanding and multimodal interaction - 🏢 Enterprise Customization : Train specialized models based on internal enterprise systems - 🌐 Domain Adaptation : Fine-tune models for specific domains healthcare, finance, etc. - 🔬 Algorithm Innovation : Develop new training techniques building on Mano-Action | Feature | Mano-P | OpenClaw | Manus | Traditional RPA | |---|---|---|---|---| | Model Source | ✅ Built-in edge model | | | ❌ No model rule-based engine | | Data Security | ✅ Local execution | | | ✅ Can be local | | Control Method | ✅ Pure visual | | ❌ HTML parsing+CLI | ❌ System API | | Use Scenarios | ✅ All-type GUI desktop/Web/3D | ✅ Multi-type apps | | | | Long Task Plan | ✅ Autonomous planning | ✅ Autonomous planning | ✅ Visual flow orchestration | ❌ Needs preset workflows | | Response Speed | ✅ Instant response | ✅ Local/cloud execution | | ✅ Instant response | | Deployment Cost | ✅ Low-cost entry | ✅ Open source & free | | ✅ Low cost | | Robustness | ✅ UI change adaptive | ✅ LLM adaptive | | ❌ UI change needs reconfig | 1. Edge Large Model + Flexible Deployment - 4B model runs directly on Mac M4 chip + 32GB RAM - Large parameter models 72B supported via computing stick - No API key configuration needed, ready out-of-the-box - Significant advantage over OpenClaw requires user model configuration and Manus cloud calls 2. Universal Visual Understanding - Pure visual GUI interaction, not limited to browsers and web apps - Broader support than OpenClaw CDP protocol mainly for browsers and Manus web apps only - Supports desktop software, 3D applications, professional tools, and non-standard GUIs 3. Offline Long-Task Autonomous Planning - Fully offline reasoning for complex business processes - Autonomous decision-making and error correction without internet connection - Unique advantage over Manus cloud latency and traditional RPA needs preset workflows 4. Integrated Hardware Deployment - Model + computing stick integrated solution, plug-and-play - Lowers technical barrier compared to OpenClaw open-source & free but requires self-deployment - Multiple deployment forms direct Mac install / plug-and-play compute stick , rapid launch Mano-AFK — an autonomous full-cycle app builder that turns a single natural-language sentence into a deployed, tested, and bug-fixed application. Its E2E testing stage runs by default with Mano-P as the local backend screenshots and task descriptions stay on-device ; it can also switch to cloud mode driven by Anthropic's Claude CUA requires ANTHROPIC API KEY ; third-party service with no local alternative . Mano-AFK is a concrete application scenario of Mano-P in real-world software engineering pipelines. - GitHub: github.com/Mininglamp-AI/mano-afk https://github.com/Mininglamp-AI/mano-afk - ClawHub: clawhub.ai/hanningwang/mano-afk https://clawhub.ai/hanningwang/mano-afk CUA Benchmark — Mano-CUA-4B on MacBook Pro M5 16GB The suite evaluates 100 tasks across 5 web applications that were themselves built autonomously by Mano-AFK: TripSplit expense splitting , md-wechat Markdown → WeChat formatter , OMS order management , Family Ledger household bookkeeping , and Life Dashboard personal widgets . Each app ships in two variants — a golden build bug-free, expected verdict PASS, 76 tasks and a buggy build with specific UI/logic defects injected expected verdict FAIL, 24 tasks . Accuracy is defined as the share of tasks where the judge's verdict matches the expected label; each project contributes 15–16 golden tasks and 4–5 bug-injection tasks. | Configuration | Accuracy | Avg Steps | Prefill Speed | Avg Tokens/Step | |---|---|---|---|---| | W8A16 | 58.0% | 6.1 | ~1,253 tok/s | 3,389 | | W8A8 Cider https://github.com/Mininglamp-AI/cider | 54.0% | 6.93 | ~1,453 tok/s | 3,104 | Metrics: Accuracy — judge verdict matches expected PASS/FAIL; Steps — actions taken per task; Prefill Speed — token throughput during the prefill phase; Tokens/Step — prompt + generation tokens combined. Note on W8A8 on this hardware: The W8A8 row runs via Cider https://github.com/Mininglamp-AI/cider for INT8 activation quantization. W8A8 accelerates prefill through INT8 TensorOps, but it must hold both the original and the INT8 weights in memory simultaneously — roughly doubling weight footprint. On a 16 GB unified-memory device, the added pressure can trigger swapping that offsets the prefill gain, which is why W8A8 here lands slightly behind W8A16 on accuracy. At least 4 GB of free memory beyond the model size is recommended to see W8A8's full benefit. - Benchmark methodology https://github.com/Mininglamp-AI/mano-afk/tree/master/benchmark — full test design, task schema, and bug catalog Mano-P is based on the following research work: 1. Mano Series Model Foundation Paper @article{mano-2025, title={Mano Technical Report}, author={Tianyu Fu, Anyang Su, Chenxu Zhao, Hanning Wang, Minghui Wu, Zhe Yu, Fei Hu, Mingjia Shi, Wei Dong, Jiayao Wang, Yuyang Chen, Ruiyang Yu, Siran Peng, Menglin Li, Nan Huang, Haitian Wei, Jiawei Yu, Yi Xin, Xilin Zhao, Kai Gu, Ping Jiang, Sifan Zhou, Shuo Wang}, journal={arXiv preprint arXiv:2509.17336}, year={2025}, url={https://arxiv.org/abs/2509.17336} } 2. WebRetriever Benchmark @article{webretriever-2026, title={WebRetriever: A Large-Scale Comprehensive Benchmark for Efficient Web Agent Evaluation}, author={Wei Dong and Tianyu Fu and Zhe Yu and Hanning Wang and Anyang Su and Zhizhou Fang and Yuyang Chen and Shuo Wang and Minghui Wu and Ping Jiang and Zhen Lei and Chenxu Zhao}, year={2026}, note={To be published}, url={https://github.com/hhhhhhalf/WebRetriever} } We welcome collaboration with academia: - 🔬 Dataset Contribution : Provide new GUI task datasets - 🤝 Joint Research : Collaborate on edge deployment, quantization optimization, GUI understanding, etc. - 📚 Benchmarking : Test Mano-P on new evaluation sets For academic collaboration inquiries, please contact: model@mininglamp.com mailto:model@mininglamp.com 🤖 What is Mano-P? Mano-P is an open-source GUI-VLA Vision-Language-Action agent designed to run locally on Apple Silicon edge devices. It uses pure visual understanding to automate desktop GUI operations across platforms. ⚖️ How does Mano-P compare to Claude Computer Use? Performance Comparison: - OSWorld all models : Claude Sonnet 4.6 72.1% vs Mano-CUA 1.1 58.2% - WebRetriever Protocol I: Mano-CUA 1.1 41.7 NavEval vs Claude 4.5 Computer Use 31.3 Key Difference: - ✅ Mano-P runs entirely on-device , no data leaves the machine - ⚠️ Claude Computer Use requires cloud API calls Use Case: Mano-P is particularly suitable for high-security scenarios . 🔌 Can Mano-P run without internet? Yes In local mode, all model inference runs on the Apple M4 device. ✅ No screenshots or task descriptions are sent to external servers. 💻 What hardware do I need? Minimum Requirements: - Mac mini or MacBook - Apple M4 chip - 32GB RAM Alternative: - Any Mac + Mano-P computing stick connected via USB 4.0+ 📌 We plan to support more devices in the future. 🔒 Is my data safe? Local Mode: ✅ All processing happens on-device Cloud Mode: - ⚠️ Only screenshots and task descriptions sent to mano.mininglamp.com - ✅ No local files, clipboard contents, or credentials accessed Transparency: Full client is open-source https://github.com/Mininglamp-AI/mano-skill for audit We welcome community contributions If you want to contribute to the project: 1. Fork this repository 2. Create your feature branch git checkout -b feature/AmazingFeature 3. Commit your changes git commit -m 'Add some AmazingFeature' 4. Push to the branch git push origin feature/AmazingFeature 5. Open a Pull Request - 🐛 Bug fixes and issue reporting - 📝 Documentation improvements and translations - 💡 New feature suggestions and implementations - 🧪 Test cases and benchmarking - 🎨 Application scenarios and demo contributions This project is licensed under the Apache License 2.0 https://github.com/Mininglamp-AI/Mano-P/blob/main/LICENSE . License Highlights: - ✅ Commercial use - ✅ Modification and distribution - ✅ Patent grant - ⚠️ Must retain copyright notice - ⚠️ Must state changes - 📧 Email: model@mininglamp.com mailto:model@mininglamp.com - 🏠 Website: https://github.com/Mininglamp-AI/Mano-P https://github.com/Mininglamp-AI/Mano-P - 💬 Community: To be added - 🐛 GitHub Issues: https://github.com/Mininglamp-AI/Mano-P/issues https://github.com/Mininglamp-AI/Mano-P/issues Thanks to all developers and researchers who contributed to this project. Special Thanks: - Mano project team for providing the technical foundation - DeepMiner platform for deep integration support - Edge computing hardware partners - Open source community contributors <sub Built with ❤️ by the Mano-P Team</sub