Client-side agent w/a server only for proxying LLM calls, 3 WebAssembly libraries: - ImageMagick: format & file manipulation - Skia: drawing, vector graphics & filters - PDFium: PDF manipulation Exploring: - Context Management: keep LLM context small and cost-effective managing large number of images - Structured Data Tools: record notes, output JSON, and create custom JSON schema-backed data stores - Vision Capabilities: compare vision capabilities of different model providers. open-weights models via OpenRouter, OpenAI, Google Gemini, and Anthropic

Cool! Context management: When you say „remove access to the image“ how do you do that? Do you still keep the „intent / edit history“ in context or start a new session?

I rewrite the context on each turn removing all exposed images and leaving a mark saying "you saw image X here and recorded the following: ..."