Marginalia

A document research agent that keeps files locally, produces cited answers and runs as a desktop app or a self-hosted Docker backend.

Marginalia is a research agent for people whose papers, notes and working files are spread across a mixed document library. It reads relevant parts of the original files before writing answers with citations.

Cited research: Ask for quick lookups or deeper investigations that follow related entries and revise searches before producing an answer or report.
Mixed document libraries: Search text, Markdown, PDFs, Word documents, spreadsheets, logs and archives; image descriptions and scanned-PDF OCR require a configured vision model.
Library organization: Keep files in readable folders by default, with catalogs, tags, metadata and relationships between entries.
Investigation memory: Save research notes that later conversations can recall alongside the source material.
Agent access: An MCP server gives external agents such as Claude Desktop access to library search, source reading and document tools.

Text search is the default. Optional semantic search requires an embedding provider, whose default is Alibaba Cloud Model Studio's text-embedding-v4; reranking is also optional. Local file storage doesn't make those external embedding requests local.

Marginalia is open source under AGPL-3.0. Desktop apps run on Windows, macOS and Linux, including Intel and Apple Silicon Macs. The desktop app and CLI share a library, and the backend can run on a self-hosted server through Docker.