Our reading system: corpus-core + two MCP servers
In the daily-feeds post we called those feeds “the public face of a larger reading system.” Here it is — three new open-source repositories that let an AI assistant actually read alongside us, over the Model Context Protocol (MCP).
corpus-core — the shared engine. Embeddings for semantic search, plain text search, a section-aware chunker, a job registry, throttled fetching, and a generic MCP-server scaffold. It carries no project-specific logic; the two servers below are both built on it.
arxiv-radar-mcp — a local MCP server over the daily arXiv feeds. Search the abstracts — semantically or by text, across science domains — and when a paper looks promising, pull its full text on demand: the server fetches and indexes it, so you can then search inside it (“found in the Methods of paper X”). Abstracts stay in memory; full texts are added to the corpus only when you ask.
lab-corpus-mcp — the broader research workspace. It ingests any literature, not just arXiv — PDFs and presentation slides — parsed with MinerU and made searchable through the same embedding stack. Lecture videos and slide decks are on the roadmap.
All three were born for our own needs: keeping one independent researcher (and an AI partner) on top of a fast-moving, cross-disciplinary literature. We use them every day, keep improving them, and we’re sharing the whole stack with the community. Open source (MIT) — and we’re open to collaboration.