What is EdgeClaw
EdgeClaw is an open-source AI agent framework launched by ModelBest in collaboration with Tsinghua University, OpenBMB, and others, focused on device-cloud collaboration, local deployment, and security and controllability. The framework pioneers a three-tier data security protocol (S1/S2/S3), allowing public information to go to the cloud, sensitive data to be desensitized, and private content to be locked down locally; combined with a dual-track memory mechanism and the GuardAgent protocol, it achieves full-process data control. EdgeClaw processes simple tasks locally with zero Token consumption, and intelligently routes complex tasks to the cloud. It has been adapted forPinecone Pi, NVIDIA DGX Spark, and other hardware, and has launched the EdgeClaw Box ready-to-use hardware, enabling non-technical users to easily have a secure and controllable AI digital employee.
Main features of EdgeClaw
- Agent execution framework: Optimized based on the OpenClaw architecture, it supports AI agents running securely on local devices without relying on the cloud.
- Three-tier data security protocol: Through S1 (public information to the cloud), S2 (sensitive information desensitization), and S3 (private data locked down locally) tiered protection, it ensures data security and controllability.
- Device-cloud collaborative computing: Simple tasks are processed locally (zero Token consumption), and complex tasks are automatically routed to the cloud, balancing performance and cost.
- Dual-track memory mechanism: MEMORY.md (desensitized version for the cloud) and MEMORY-FULL.md (full version limited to local only), enabling tiered data management.
- Local model support: Integrates the ModelBest MiniCPM on-device model, and can still run in offline environments, ensuring privacy and availability.
- GuardAgent protocol: Six-stage lifecycle hooks provide full-process control over sensitive data.
Key information and usage requirements of EdgeClaw
- Developer: ModelBest, Tsinghua University THUNLP, Renmin University of China, AI9Stars, OpenBMB
- License: MIT open-source license
- Tech stack：TypeScript（83.3%）、Swift（12.7%）、Kotlin（1.8%）
- Core positioning: A device-cloud collaborative AI framework that supports both models and agents being deployed locally
- Hardware requirements: Supports mainstream edge devices such as Pinecone Pi, NVIDIA DGX Spark, and Apple Mac Mini, with optional EdgeClaw Box hardware available ready to use.
- Software environment: Requires configuring a local runtime environment (simplifying the deployment process compared with OpenClaw), and supports integrating on-device models such as MiniCPM for offline operation.
Core advantages of EdgeClaw
- Secure and controllable: Pioneers a three-tier data security protocol (S1/S2/S3), enabling public information to go to the cloud, sensitive data to be desensitized, and private content to be locked down locally, combined with GuardAgent six-stage lifecycle control, solving data leakage risks from the source.
- Device-cloud collaboration: Simple tasks consume zero Tokens locally, and complex tasks are intelligently routed to the cloud, balancing privacy protection and performance needs, with marginal costs approaching zero.
- Local-first: Integrates the MiniCPM on-device model, usable offline, with no need to worry about network fluctuations or service interruptions, truly achieving autonomous control of data sovereignty.
- Ready to use: The framework simplifies the deployment process, has been adapted for mainstream edge hardware, and has launched the EdgeClaw Box hardware product.
- Cost advantage: Supports local execution of daily tasks, reducing costs by several orders of magnitude compared to pure cloud solutions, and is especially suitable for high-frequency, high-privacy scenarios.
How to Use EdgeClaw
- Environment Preparation: Install the EdgeClaw framework and configure the local runtime environment. Choose an adaptation solution based on hardware: edge devices such as Pinecone Pi, NVIDIA DGX Spark, Mac Mini, etc., or directly purchase EdgeClaw Box hardware for out-of-the-box use.
- Model Deployment: Download and integrate the ModelBest MiniCPM on-device model to ensure local offline operation capability. Configure cloud model interfaces as needed to handle complex tasks.
- Security Configuration: Enable the three-layer data security protocol: set S1/S2/S3 data classification rules, configure GuardAgent lifecycle hooks, and establish a dual-track memory mechanism (MEMORY.md desensitized version, MEMORY-FULL.md full version).
- Startup and Operation: Start the local Gateway and maintain intranet access (avoid direct public network exposure). Invoke the agent through the API or interface, and the system automatically determines task complexity, handling it locally or coordinating with the cloud.
EdgeClaw Project Address
- GitHub Repository：https://github.com/OpenBMB/EdgeClaw
Comparison of EdgeClaw with Similar Competitors
| Product | Developer | Core Difference | Advantages | Disadvantages |
|---|---|---|---|---|
| EdgeClaw | ModelBest/Tsinghua | Device-cloud collaboration, three-layer security protocol | Data classification protection, local-first, low cost, usable offline | Ecosystem maturity needs improvement |
| OpenClaw | Anthropic | Cloud-first remote control framework | Rich features, complete ecosystem | Data must be uploaded to the cloud, posing leakage risks |
| Tencent QClaw | Tencent | WeChat ecosystem integration, minimalist encapsulation | WeChat Mini Program entry point, low barrier to entry, suitable for domestic users | Depends on the WeChat ecosystem, data privacy depends on Tencent |
| JVS Claw | Alibaba | Cloud AI Agent platform | Sufficient free quota, strong cloud computing power | Pure cloud solution, data leaves the local environment |
EdgeClaw Application Scenarios
- Personal Privacy Protection: Supports local processing of private data such as chat records, photos, and documents, keeping sensitive information off the cloud and preventing cloud leakage risks.
- Enterprise Data Security: In industries such as finance, healthcare, and law, it processes confidential information such as customer data, medical records, and contracts, meets compliance requirements, and enables autonomous control of data sovereignty.
- One-Person Company (OPC): Entrepreneurs use EdgeClaw Box to build a personal digital employee team, completing daily tasks such as document processing, schedule management, and content creation at low cost.
- Offline Environment Operations: In field surveys, remote areas, and scenarios with unstable networks, it relies on on-device models to independently complete data analysis, report generation, and other work.