Marginalia

Marginalia is a local-first AI research agent for people whose research papers, notes and working documents are scattered across different formats. It checks relevant passages in the original files before writing answers with citations. Your library stays in readable folders by default.

It handles PDFs, Markdown, Word documents, spreadsheets, logs and archives. Image descriptions and scanned-PDF OCR require a vision model. Folders, tags, catalogs and relationships between entries help organize the material, so you can compare papers, trace an incident across logs or ask for a report that cites specific spreadsheet rows.

Quick mode handles short lookups; Deep mode spends longer investigating. The agent can follow related documents, revise its search and reuse saved investigation notes in later conversations. Text search works by default, with optional semantic search and reranking for finding and selecting evidence.

The open source app uses the AGPL-3.0 license and runs on Windows, macOS and Linux, including Intel and Apple Silicon Macs. You can also self-host its backend with Docker. The desktop app, command-line interface and MCP server share the same library, and MCP lets external agents such as Claude Desktop search and read its sources.

Local storage doesn't mean every AI request stays on your machine. Marginalia supports OpenAI, Anthropic and OpenAI-compatible endpoints; cloud providers receive the content sent in model requests. Optional embeddings use a separately configured provider, with Alibaba Cloud Model Studio's text-embedding-v4 as the default.