HashAgent deploys browser-local agent execution over WebGPU with URL share links

HashAgent encodes model weights references, prompt templates, and tool definitions into a compact URL fragment that a recipient opens to instantiate the identical agent instance locally. The implementation relies on WebGPU compute shaders for matrix operations and the Transformers.js runtime for tokenization and sampling. No external API keys or hosted endpoints are required once the page loads.

WebGPU adoption data from Chrome Platform Status shows stable compute shader support across 78 percent of desktop Chrome sessions as of Q3 2024, yet measured FP16 throughput on consumer GPUs remains 3-8x below dedicated CUDA runtimes on identical models. HashAgent therefore targets sub-7B parameter models where browser latency stays under 120 ms per token. The project omits any published benchmark suite comparing output quality or memory footprint against local alternatives such as Ollama or llama.cpp with Vulkan backends.

Security analysis is absent from the deployment. A shared URL can embed arbitrary system prompts and tool schemas that execute without user review once the page loads, creating an attack surface equivalent to cross-site scripting but delivered through model instructions. Related incidents documented in the 2024 OWASP LLM Top 10 list already record prompt-injection vectors achieving remote code execution when tool-calling interfaces lack sandboxing.

Operational deployment requires Chrome 113+ or Edge 113+ with hardware acceleration enabled; fallback paths to WebGL or CPU are not implemented. Future iterations will need explicit permission prompts for tool execution and reproducible benchmark reports before wider adoption can be assessed.