Dataroom

Dataroom hands a query to a small, self-hosted language model running through llama.cpp inside a headless coding-agent harness, which loops through searching, reading, deduplicating, writing, and verifying sources until it has produced a structured dataroom directory on disk: a topics folder, a sources folder, a data folder, and a cited summary file. Because gathering and organizing sources is mostly tool-calling rather than deep reasoning, the project argues a disciplined local model can do it well and, running on your own GPU at near-zero marginal token cost, can keep going for hours rather than stopping when a metered budget runs out.

The resulting zip is meant to be unzipped directly into a frontier model's context for the more expensive second stage of a task, such as implementation, rather than read by a person the way a long research report would be. An async job spins up the agent per submitted query, resuming the same per-working-directory session across turns; a premature done signal from the agent is rejected until the dataroom holds enough substantive sourced files, every sub-question is closed, and a summary exists, with turn, time, and search-call counts only as hard backstops. Everything runs locally except the web search and page fetches the agent chooses to make.