Mano-P: a GUI-VLA agent that runs entirely on an M4 Mac
Mano-P: an open-source GUI-VLA agent for edge devices, ranked #1 on OSWorld (specialized, 58.2%). Runs locally on an Apple M4 Mac mini/MacBook or via a compute stick for purely vision-driven, cross-platform GUI automation with complex multi-step task planning — no data leaves your device.
At a glance
- What is it?
- Mininglamp's Mano-P is an Apache-2.0 GUI-VLA project aimed at on-device desktop automation. The repository currently ships documentation and the Mano-CUA Skills phase; the local model SDK and training code are announced as later phases.
- Who is it for?
- Mano-P is worth tracking if you want GUI automation where screenshots and task data never leave an M4 Mac mini or MacBook with 32GB of RAM, and the Mano-CUA Skills phase is the part you can act on today. It is not the right choice if you need a fully open training pipeline now, since the repository states that local models, the SDK, and the quantization methods arrive in later phases, and no releases have been published.
- Can I use it commercially?
- Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 144 days ago.
- What is it written in?
- GitHub does not report a main language for this repository.
Answers come from the project's GitHub data, last synced on November 14, 2026, and from our analysis. They are not legal advice.
Star history
- Mininglamp-AI/Mano-P2,849 ★ · +60
| Repository | 2026-09-25 | 2026-11-14 |
|---|---|---|
| Mininglamp-AI/Mano-P | 2.8k | 2.8k |
60 Stars in the last 50 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
The problem Mano-P targets: GUI automation without sending screenshots to a server
Most computer-use agents work by shipping a screenshot to a hosted model and getting back a click coordinate. That is fine for a demo and awkward for anything involving internal dashboards, customer records, or regulated data. Mano-P is built around the opposite arrangement. The README describes it as a GUI-VLA agent project designed specifically for edge devices, where inference runs locally on an Apple M4 chip with 32GB of RAM or more, on a Mac mini or MacBook, and the project states that all CUA operations execute on the local machine and are not uploaded to external servers.
The intended audience is narrower than a general automation library. The README splits the project into three phases for three developer groups: agent enthusiasts building CUA workflows on top of Mano-CUA Skills, security-conscious developers who want to run the GUI-VLA model locally and build their own skills and tools, and developers who want to train their own on-device models using Mano-P's training, pruning and quantization methods. Only the first phase is described as open-sourced so far. That distinction matters more than any benchmark number on the page.
How the Mano-P architecture is described: a model, a skills layer, and a quantization SDK
The repository layout is thin. The top level holds .gitignore, LICENSE, README.md, README_CN.md and a pics directory. There is no source tree to read, so the architecture has to be taken from the README's own description rather than from code.
What the README sets out is a three-part stack. At the bottom sits a GUI-VLA model, Mano-CUA 2.0 4B, published on Hugging Face and ModelScope, which is the component that runs inference on the Mac. Above it sit Mano-CUA Skills, the piece described as open-sourced first, which let an agent framework call GUI capabilities as tools. Alongside both is Cider, an inference SDK that the README says adds W8A8 and W4A8 activation-quantization primitives that MLX lacks natively, and which the project says works with any MLX model, not only Mano-P.
The training story is described as a three-stage progression: supervised fine-tuning, then offline reinforcement learning, then online reinforcement learning, wrapped in a think-act-verify loop and what the README calls Mano-Action bidirectional self-reinforcement learning. The edge version adds mixed-precision quantization and visual token pruning. None of that is visible in the repository yet, so treat the description as a roadmap rather than as something you can inspect.
Installing Mano-P and running a first task
The README does not contain install commands. It points to Hugging Face and ModelScope for the weights and to a Feishu wiki link for the benchmark overview, and it says detailed instructions for both deployment methods, direct on an M4 Mac and via a compute stick on USB 4.0 or higher, will be released in the near future. Because no install steps are given, the honest starting point is to clone the repository and read what is actually there.
git clone https://github.com/Mininglamp-AI/Mano-P.git
cd Mano-P
lsAfter that listing you should see the entries the repository documents: LICENSE, README.md, README_CN.md, pics and .gitignore. There is no setup script, no requirements file and no entry point in that listing, which tells you the runnable Skills code is not in this tree as of the last push on 2026-06-25.
The weights are published separately. The README links the Mano-CUA 2.0 4B model on Hugging Face and on ModelScope, so a first real use begins by pulling the model from one of those two hosts rather than from the repository itself. The second deployment path in the README is a compute stick connected over USB 4.0 or higher, for machines that do not meet the M4 and 32GB requirement. The README states that instructions for both paths are coming; it does not give the commands.
Where Mano-P is the wrong tool
The hardware floor is the first constraint. The README names an Apple M4 chip with 32GB of RAM or more, or a compute stick on USB 4.0 or higher. A Linux workstation with a discrete GPU is not in that list, and neither is an older Intel Mac. If your fleet is mixed, Mano-P is not a drop-in automation layer.
The second constraint is what is actually in the repository. The README describes a phased release and states that the local-side models and SDK components come in phase two and the training, pruning and quantization methods in phase three. The repository contains documentation and images. If you need to read the training code or audit the quantization implementation before trusting it with internal screenshots, that code is not here to read.
The third is the absence of releases. The repository has no published GitHub releases, so there is no tagged version to pin and no changelog to diff against. Combined with a last push on 2026-06-25, that means the project is documented and announced but not yet packaged for consumption. A team that needs a stable dependency today should look elsewhere and come back when phase two lands.
Mano-P compared with OpenCUA and the hosted computer-use agents
The README positions Mano-P against two groups. The first is open GUI agents: it reports Mano-CUA 1.1 at 58.2% on OSWorld, first among specialized GUI agent models, ahead of opencua-72b at 45.0%. The difference in approach is parameter count versus placement. A 72B model in that comparison is not something you run on a laptop; Mano-P's edge story rests on a 4B model plus quantization and token pruning so that inference fits on a Mac. That is a deliberate trade of raw capacity for locality.
The second group is the hosted computer-use models. The README reports 41.7 NavEval on WebRetriever Protocol I for Mano-CUA 1.1, against Gemini 2.5 Pro Computer Use at 40.9 and Claude 4.5 Computer Use at 31.3. Those numbers are the project's own reported results, and the comparison is not like for like: the hosted models are general assistants with broad API surfaces, while Mano-P is a specialist that has to run on your desk. If your task involves data you are comfortable sending to a vendor, the hosted route removes the hardware requirement entirely. Mano-P's argument only holds when locality is the requirement, not a preference.
Maintenance status, licence and upgrade cost
The repository is not archived, and the last push was on 2026-06-25. That is roughly three months before today, so it is recent enough that the project is not dormant, but there is no release history to read, no tagged version, and the README describes phases that have not shipped. Calling this a stable dependency would overstate what the repository shows.
Upgrade cost is hard to estimate for the same reason. Without releases there is no version boundary to plan around, so an early adopter is tracking the main branch of a documentation repository and waiting for phase two to appear. The hardware side is a capital cost rather than an upgrade cost: an M4 Mac mini or MacBook with 32GB of RAM, or a USB 4.0 compute stick.
The licence is Apache-2.0, both in the LICENSE file at the repository root and on the badge. Apache-2.0 is permissive and includes an explicit patent grant, which is generally the friendlier option for commercial integration than a copyleft licence. That applies to the repository contents. The model weights are hosted on Hugging Face and ModelScope and the README does not state their licence terms in the text available here, so check the model card on whichever host you pull from before shipping anything. This is a description of what the files say, not legal advice.
Editorial conclusion
Mano-P is worth tracking if you want GUI automation where screenshots and task data never leave an M4 Mac mini or MacBook with 32GB of RAM, and the Mano-CUA Skills phase is the part you can act on today. It is not the right choice if you need a fully open training pipeline now, since the repository states that local models, the SDK, and the quantization methods arrive in later phases, and no releases have been published. Before adopting anything, confirm which phase actually contains the component you need, check that the Hugging Face or ModelScope weights you intend to use match the model named in the README, and verify the licence file in the repository root rather than the badge.
Frequently asked questions
What hardware does Mano-P need to run locally?
The README states that inference runs locally on an Apple M4 chip with 32GB of RAM or more, on a Mac mini or MacBook. It also describes a second deployment path using a compute stick connected over a USB 4.0 port or higher. Instructions for both paths are described as coming in the near future.
Does Mano-P send screenshots or task data to a server?
The README states that all CUA operations execute entirely on the local Mac mini and are not uploaded to external servers, and that all screenshots and task data stay on-device. That is the project's stated design goal for the edge version.
What is Mano-P 2.0 and what is in the repository right now?
The README describes Mano-P as a GUI-VLA agent project for edge devices, released in three phases. The first phase, Mano-CUA Skills, is the one described as open-sourced; the local models and SDK come in phase two and the training and quantization methods in phase three. The repository top level currently contains .gitignore, LICENSE, README.md, README_CN.md and pics.
Where do I download the Mano-P model weights?
The README links the Mano-CUA 2.0 4B weights on Hugging Face and on ModelScope, and the repository itself does not host them. The README does not give download commands, so use the model card on whichever of those two hosts you choose.
What does the P in Mano-P stand for?
The README states that Mano means hand in Spanish and that the P stands for Private. The project frames this as the idea that individuals and organizations can build their own private AI.
Official sources
Where developers are discussing it
Posts on Hacker News, dev.to and Lobsters that link to this repository, found by our scan of those communities. 1 in the last 30 days; newest first.
- Vision-Only Desktop Agents vs. API-Based: Running the Same Task Both WaysHacker News · Oct 21, 2026 · 1 points · 0 comments
- Mano-P: Setting Up an On-Device GUI Agent Without Cloud DependenciesHacker News · Oct 14, 2026 · 1 points · 0 comments
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/mininglamp-ai-mano-p)Star history chart for your README
[](https://hysenlabs.com/projects/mininglamp-ai-mano-p)Was this useful?