EdgeClaw Launches: A Secure and Efficient Device-Cloud Collaborative Agent
The wildly popular OpenClaw has made more and more developers and individual users realize how much a personal agent can do. Yet in practice, users widely raise two problems: first, personal data is uploaded by OpenClaw to cloud large models, causing some degree of personal data leakage; second, the extremely long context assembled during OpenClaw execution wastes tokens and drives up usage costs.
Today, we are officially releasing and open-sourcing EdgeClaw, a secure and efficient device-cloud collaborative agent framework jointly developed on top of OpenClaw by THUNLP, Renmin University of China, AI9Stars, ModelBest, and OpenBMB. Through three-tier security collaboration and a cost-effectiveness-aware collaboration mechanism, it addresses local data leakage and high token costs during OpenClaw use. Deployed on desktop devices such as DGXSpark and MacMini, it gives users a secure and efficient local lobster experience.
➤ GitHub link
🔗 https://github.com/Openbmb/EdgeClaw
EdgeClaw: A Secure and Efficient Device-Cloud Collaborative Agent
In today's AI agent architectures, the device side has long been neglected — all data and tasks pour into the cloud, giving rise to privacy leakage and wasted compute.
EdgeClaw uses a three-tier security collaboration mechanism for local data encryption and secure isolation, and a cost-effectiveness-aware collaboration mechanism to flexibly call models at different price points, so simple tasks use cheaper models and complex tasks use higher-end ones. Security collaboration and cost-effectiveness collaboration run in the same pipeline, coordinated through weights and a two-stage short-circuit strategy.
EdgeClaw's main functionality is implemented as an OpenClaw plugin. With device-cloud intelligent forwarding, developers can achieve "public data to the cloud, sensitive data desensitized, private data kept on-device" — frictionless device-cloud privacy protection and cost savings — without changing business logic.
Three-Tier Security Collaboration
A core innovation of EdgeClaw is its self-developed three-tier security collaboration mechanism: by inserting hooks into the OpenClaw execution flow, EdgeClaw automatically classifies every user message, tool-call parameter, and agent output into three levels by sensitivity:
S1: default mode (information is processed in the cloud), handled directly by cloud models.
S2: desensitization mode (information is processed after desensitization), automatically blurring sensitive enterprise information (for example, turning "Wang Xiaoer" into "Employee A") before sending it to the cloud.
S3: secure mode (information must be processed locally), physically isolated, with sensitive data kept entirely on-device and processed offline by pre-installed MiniCPM series models.
This article includes video content. Please view it in the original WeChat post.
For private files, EdgeClaw identifies them as S3 and processes them offline with the local model.
On top of the three-tier classification, EdgeClaw uses a dual detection engine based on a rule-based detector and a local LLM detector. The two engines can be combined and stacked, enabled flexibly per scenario through the checkpoints configuration.
Meanwhile, EdgeClaw maintains a "dual-track memory" mechanism — cloud models can only see the desensitized conversation history, and only local models can access memory containing full information, fundamentally eliminating the risk of private data leaking to third-party cloud services through the context window.
Cost-Effectiveness-Aware Collaboration
In a typical AI coding assistant workflow, most requests are file lookups, code reading, and simple Q&A — handling them with the most expensive model is hugely wasteful. Cost-effectiveness-aware collaboration uses a small local model as an LLM-as-Judge, routing requests by complexity to cloud models at different price points.
Table: Example routing model configuration
In actual tests with a typical coding assistant workflow, the cost-effectiveness-aware collaboration mechanism routed 60–80% of requests to cheaper models.
Composable Routing Pipeline
Security collaboration and cost-effectiveness collaboration run in the same pipeline, coordinated through weights and a two-stage short-circuit strategy. The pipeline is designed security-first: the security router runs first with high weight; if sensitive data is present, it short-circuits immediately, wasting no time on complexity judgment. Only after passing security (S1) does cost-effectiveness optimization start.
Future Plans
EdgeClaw will keep iterating and evolve into a software system supporting deployment on a broader range of on-device devices. The future development plan has four parts:
EdgeClaw Router. Focused on secure and efficient device-cloud collaboration, integrating more on-device hardware and models to enable more flexible and diverse local model choices.
EdgeClaw Memory. Optimize OpenClaw memory to implement task-oriented memory mechanisms, enabling agents to genuinely support users in long-term complex tasks.
EdgeClaw SkillHub. Build a Skill Hub focused on local tasks, integrating the specific skills that on-device agent services care about and bundling high-frequency skills. For mainstream skills implemented with AI models, provide on-device model alternatives to further improve efficiency.
EdgeClaw UI. Support front-end UI designs friendlier to on-device scenarios, adding performance monitoring metrics that on-device users care about, such as local GPU utilization and local token usage.
EdgeClaw focuses on building secure and efficient device-cloud collaborative agents and will remain open-source. Developers and industry partners are welcome to contribute and jointly build a "more, faster, better, cheaper" on-device agent solution.