SuperAPI AI Directory

EdgeClaw Box Experience - Open-source device-cloud collaborative AI hardware, runs Agents even when offline

Author: AI小集 · Original source: AI工具集

Recently, complaints about OpenClaw have been increasing. Not only is deployment difficult, but API token consumption is also high. When running complex automation tasks, one look at the bill makes my wallet ache.

Most critically, the Ministry of State Security has also begun warning that OpenClaw has security risks. How would we dare hand over our core work materials to it for processing?

Yesterday, ModelBest released an intriguing new product, a lobster intelligent hardware focused on safety, controllability, and out-of-the-box usability—EdgeClaw Box.

EdgeClaw Box is equipped with a new generation of device-cloud collaborative, security-first lobster jointly developed and open-sourced by ModelBest in collaboration with Tsinghua University, Renmin University of China, AI9Stars, ModelBest, and the OpenBMB community:EdgeClaw。

Official website:https://github.com/Openbmb/edgeclaw

EdgeClaw Box can be said to be the most perfect OPC digital infrastructure currently available. It allows the lobster to run safely, stably, and attentively in real work scenarios.

 

01. AI-native hardware EdgeClaw Box

 

In the past, we bought computers to do work ourselves; but EdgeClaw Box is equivalent to a one-time buyout, hiring a cyber employee that never sleeps to do the work for you. This is also the industry's first AI-native hardware form factor.

Everyone knows OpenClaw has weak spots: difficult local deployment and easy data leakage. And security is precisely EdgeClaw Box's first pursuit.

In terms of model architecture, EdgeClaw Box unprecedentedly breaks through the dimensional barrier between "device" and "cloud," becoming a true "amphibious lobster":

  • On-device (strongest defense): Built-in ModelBest full suite (including the extremely powerful on-device king MiniCPM series), can work even when offline, with data completely physically isolated.
  • On-cloud (switch at will): Freely switch between mainstream large models such as Kimi, MiniMax, and Claude.

We can connect directly via Bluetooth to EdgeClaw Box using a phone, AR glasses, or even our main computer. We open Feishu, or the dedicated APP developed by ModelBest itself, and send voice messages and assign tasks just like sending a WeChat message to a colleague.

For example, while walking on the road, we send a single instruction, and the EdgeClaw Box in the office can complete the entire workflow.

The efficient RAG is extremely smooth. All our data can be indexed locally, making it both secure and efficient.
ModelBest designed three very ingenious mechanisms in the underlying architecture:

Three-tier security mechanism

Regarding the data security everyone cares about most, EdgeClaw Box has three built-in task processing modes.

  • S1 default mode (cloud processing): Suitable for looking up information and writing ordinary copy. Information is sent directly to the cloud large model, pursuing ultimate intelligence and efficiency.
  • S2 desensitization mode (hybrid processing): I think this design is simply genius! The local model performs "automated desensitization processing" on the files we upload (for example, erasing or replacing names, amounts, and core code), and then sends the desensitized safe content to the cloud for inference, combining both security and efficiency.
  • S3 safe mode (forced local processing): All data flows remain completely local and are processed offline by the pre-installed MiniCPM model. Even if we unplug the network cable, EdgeClaw Box can still do the job beautifully.

Cloud / local dynamic collaboration

Many friends tend to pay close attention to token consumption. EdgeClaw Box uses dynamic cloud/local collaboration to genuinely help us save money. For simple tasks like organizing recordings and basic email replies, the system automatically calls local models to handle them, with zero latency and zero API token cost.

For complex tasks such as in-depth industry financial report analysis and long-cycle code development, EdgeClaw Box seamlessly switches to cloud-based large models. It perfectly balances local security and low cost with the cloud's strong computing power and high intelligence.

Dual-layer Skill architecture

EdgeClaw Box launches with multiple premium Skills:

  • General Skill: lets EdgeClaw Box take over miscellaneous work, such as generating meeting minutes, content creation, replying to emails, and more.
  • Professional Skill: directly empowers core business to make money. For example, with the investment analysis Skill, EdgeClaw Box can automatically monitor the market 24 hours a day, help us analyze hundreds of pages of fully English-language listed company financial reports, and distill them into summaries to send over.

 

02. Some sharing

 

EdgeClaw is truly our own AI operating system.

Thanks to ModelBest and several top universities, we can not only buy an out-of-the-box EdgeClaw Box, but also install the EdgeClaw system directly on our own computers or servers, giving us an efficient and secure AI assistant.

The future common sense of business is being rewritten. For a company to expand, it only needs to upgrade your OPC digital infrastructure and buy a few more EdgeClaw Boxes.

Technology is, in an extremely violent yet incomparably fair way, erasing the resource gap between huge corporations and super individuals.

The emergence of EdgeClaw Box has turned "one person is a team" from a motivational saying into a truly realized reality.

Original link:ModelBest releases the end-cloud amphibious lobster, which can still work even when the network is down