Google's Android team recently released a new CLI toolchain built for AI agents — packaging SDK management, project creation, and device debugging into streamlined commands with standardized Skills and a searchable Knowledge Base. It's a clear sign: on-device agents are moving from concept to production.
On mac, our open-source project Mano-P has been working on the same frontier — enabling AI agents to run locally and operate real GUI applications on your own machine.
What Is Mano-P?
Mano-P is an open-source, on-device GUI agent built by Mininglamp Technology, designed for macOS. It's based on a VLA (Vision-Language-Action) architecture.
"Mano" means "hand" in Spanish, and "P" stands for Person & Party — our vision is that every individual and organization can create their own personalized AI, running on their own hardware.
Mano-P is purely vision-driven: it understands screen content through visual models, plans action sequences, and executes operations via native OS input (mouse clicks, keyboard strokes). No system APIs or CLI access required — it can theoretically operate any GUI application on Mac.
Key Technical Features
think-act-verify Loop
Mano-P doesn't do single-pass inference. It uses a cyclical reasoning mechanism: observe → think → act → verify → repeat. After each action, the model re-examines the screen state to confirm success before deciding the next step. This enables complex workflows spanning dozens to hundreds of steps.
Edge Optimization
To run efficiently on consumer hardware, Mano-P uses mixed-precision quantization and the GS-Pruning algorithm for visual token compression. Performance of the 4B quantized model (w4a16) on Apple M4 Pro:
- Prefill: 476 tokens/s
- Decode: 76 tokens/s
- Peak memory: 4.3 GB
A Mac mini or MacBook with an M4 chip and 32GB RAM can run Mano-P locally, with no cloud dependency.
Three-Stage Training
SFT → Offline RL → Online RL, with a bidirectional self-reinforcement framework (Text↔Action cycle consistency learning).
Benchmarks
Mano-P has achieved competitive results on public benchmarks:
| Benchmark | Mano-P (72B) | Result |
|---|---|---|
| OSWorld (Specialized) | 58.2% | #1 |
| WebRetriever Protocol I | 41.7 NavEval | #1 |
Data Privacy
In local mode, Mano-P keeps all screenshots and task descriptions on-device. Nothing leaves your machine. The full client code is open-source and auditable.
Three-Phase Open Source Plan
- Phase 1 (released): Mano-CUA Skills for agent enthusiasts and Claude Code/OpenClaw users
- Phase 2 (coming soon): Local model + SDK for developers with strict data security requirements
- Phase 3 (planned): Training methods + pruning/quantization techniques
Get Started
Install via Homebrew:
brew tap HanningWang/tap && brew install mano-cua
GitHub: github.com/Mininglamp-AI/Mano-P (Apache 2.0)
The era of on-device agents is accelerating. Google took a step on Android; we've been building for Mac. If you're interested in local AI agents, give Mano-P a try and let us know what you think in the comments.
Top comments (0)