Agent Kernel Lite: A Local AI Research Assistant Running in Browsers

Agent Kernel Lite is a browser-first local AI research assistant built with Rust/WASM, integrating features like BitNet quantized models, local paper retrieval, and a verifiable extension system, supporting direct operation on devices such as iPhones. Its core advantages include privacy protection (data processed locally), no network dependency, and near-native execution efficiency, representing the trend of AI applications shifting from cloud to edge localization.

Agent Kernel Lite is separated from the larger Agent Kernel project, aiming to allow the lightweight browser application, model stack runtime, and Rust/WASM core to evolve independently. It adopts a browser-first design where all processing is done locally to ensure user data privacy; core functions can be used without network connection; and near-native execution efficiency is achieved via WebAssembly technology.

Rust/WASM Agent Core: A WASM module written in Rust, responsible for processing conversation states, context packages, model decisions, etc. Combining Rust's memory safety and WASM's sandbox environment, it provides a reliable infrastructure.

BitNet Model Integration: Using BitNet quantization technology, it compresses model weights to 1.58 bits, significantly reducing memory usage and computational requirements, enabling large language models to run on resource-constrained devices like iPhones.

Multi-Mode Conversation: Supports three modes: Chat (daily Q&A), Think (deep reasoning), and Deep (combining local paper analysis).

Local Paper Retrieval: Built-in semantic search function that can retrieve metadata and vector packages of downloaded papers, maintains persistent context of selected papers, suitable for building personal knowledge bases.

Extension System: Divided into in-app extensions (enabled/disabled by users, model outputs are not directly executed) and browser validator extensions (independently verify Web application asset hashes).

Session Backup: Supports exporting/importing JSON session packages, including UI settings, extension status, chat messages, etc., excluding large caches (model weights, paper packages, etc.) to maintain practicality.

The BitNet decoder kernel was optimized for the browser WASM environment, with decoding speed data as follows:

Encoder Context	Total Decoding Speed	Stable Decoding Speed
66 tokens	~368 tok/s	~408 tok/s
130 tokens	~360 tok/s	~413 tok/s
258 tokens	~275 tok/s	~334 tok/s
514 tokens	~176 tok/s	~226 tok/s

In local tests, the complete browser-worker thread path for generating 64 tokens took approximately 500 milliseconds, performing excellently.

Application Hash Verification: The status panel calculates the hash values of shell assets (index.html, js files, WASM packages, etc.), and users can verify consistency via the SHA256SUMS of published assets.

Computer Usage Bridging: Supports local bridging on the same computer (http://127.0.0.1:45731) and hosted HTTPS relay (/agent_kernel/api/relay/). The relay uses unguessable URLs and tokens, requiring a short pairing code + local approval to store authorization.

Agent Kernel Lite is suitable for:

1. Privacy-sensitive research (processing sensitive documents in fields like law and medicine);
2. Network-constrained environments (airplanes, remote areas, etc.);
3. Personalized knowledge management (building a private research assistant);
4. Mobile device AI applications (running on smartphones via BitNet quantization).

It represents an important trend of AI applications shifting from cloud-centralized to edge-localized.

Agent Kernel Lite is an open-source project that combines technical depth and practicality. By integrating Rust/WASM high performance, BitNet efficient quantization, and an extension architecture, it sets a new standard for local AI assistants. For developers focusing on privacy protection, edge computing, and AI innovation, it is a project worth in-depth research and contribution.