I’m Yash, the team of one behind Marginalia. It puts your notes beside the web page you are reading. Select a passage, keep it, write a note or park the thread for later. Your saved reading collects in a local library with source attribution and plain-file export. The difference I care about is the relationship between the source, your note and the reply. The page stays as its author wrote it. Your note is senior to the model’s answer. Ask can request a definition, an example or a step-by-step explanation through Codex. You review the outgoing text and recipient before it sends. For the GPT-6 Astra Challenge, I built with Astra through Codex, used Astra and ChatGPT Pro, and used Claude Code for integration and review. The build meant turning reading interactions into bounded tasks and checking the result against the code. Real asks require the optional reader-authorized mode, where solver confinement is requested but not observed; no recorded real run is in the evidence folder yet. Windows install is proven, while macOS and Linux have command-level CI evidence. PDF support is still ahead. The research whitepaper (https://yashgurbani.github.io/ma...) explains the design and its research lineage. Credits name the prior art and tools; third-party notices record dependency licenses. The build scope separates mapped work from evidence gaps. Library search and “Related saved passages” are available in the local library page. Code and automated tests support this claim; native browser acceptance has not been recorded.

Marginalia

A personalized, agentic margin for the web

A personalized, agentic margin for the web