Mano-P
Editor's PickOpen-source local GUI-VLA computer-use agent by Mininglamp — a 4B model that sees your screen and operates your Mac fully offline, #1 specialized agent on OSWorld.
Try it now
GPT-4oWhat is Mano-P?
Mano-P is an open-source GUI-VLA (Vision-Language-Action) agent from Mininglamp Technology, designed to run locally on Apple-Silicon edge devices. Using pure screen understanding, it automates cross-platform desktop GUI tasks the way a human would — looking at the screen and clicking, typing and navigating — with no cloud API, so screenshots and task data never leave the device. Its 72B model scored 58.2% on OSWorld, ranking first among specialized GUI-agent models (leading the runner-up by 13.2 points), and its WebRetriever NavEval score of 41.7 tops Gemini 2.5 Pro and Claude 4.5 Computer Use. Through distillation and quantization it is compressed to a 4B model that runs smoothly on an M4 Mac (about 4.3GB peak memory, 476 tokens/s prefill; w8A8 Cider quantization adds another 12.7%). Released under Apache 2.0 for commercial use, it offers a brew-installable CLI (mano-cua), a Python SDK (mano-client) and an AI-agent skill (mano-skill), with Mano-CUA Skills being open-sourced in phases in version 2.0. Compared with Claude Computer Use it is fully local, offline, runs 24/7 with no per-call cost and is fully auditable.
Pros
- Strong privacy: completely local execution with nothing uploaded
- State-of-the-art OSWorld result among specialized GUI agents
- A 4B variant runs on a consumer MacBook instead of requiring a data center
- Apache 2.0 allows unrestricted commercial use and code auditing
- CLI, SDK and agent-skill integrations plus no marginal API cost for long tasks
Cons
- Best experience targets Apple Silicon (M-series) edge devices
- The full 72B model still needs substantial hardware
- Setup and deployment are technical, aimed at developers
- GUI agents still fail on unfamiliar or rapidly changing interfaces
- Young ecosystem compared with large cloud computer-use products
Capabilities
Compare with Alternatives
| Feature | Mano-P | PGVector | Semantic Kernel | Warp |
|---|---|---|---|---|
| Rating | 4.6 | 4.1 | 4.8 | 4.8 |
| Pricing | Free | Free | Free | Free |
| Starting From | — | Free | Free | Free |
| API | — | — | — | — |
| Free Tier | — | — | — | — |
| Mobile App | — | — | — | — |
Who is it for
Computer-Use Developers
Build local desktop-automation agents with a VLA model via CLI or Python SDK.
Privacy/Regulated Automation
Automate repetitive desktop workflows offline so sensitive screens never leave the machine.
RPA & Ops Teams
Run unattended, 24/7 UI automation without per-call cloud fees.
Edge / On-Device Builders
Ship an on-device AI operator that runs on customer Apple hardware.
VLA Researchers
Study and fine-tune an open, benchmark-leading vision-language-action model.