MaryCache/imozuru

If you already use Claude Code and want research to end at primary sources rather than a tidy summary, imozuru is worth a look. It is a skill, not a separate app, and it deliberately stops before interpretation.

The problem

You ask an assistant to research an unfamiliar field, and you get a short summary built from the first few search results. It reads fast, but the details you need for a decision are often gone, and because you do not yet know the field's vocabulary, your search terms cannot get you past your first guesses. imozuru changes the middle of that workflow: it first maps the terminology, the people and the places where knowledge collects, then follows citations, links and authors until it has saved the originals. You end up with a folder of source files, a map of the field and a page of quotations tied to each claim.

What's different

The skill writes out your question and 3–6 claims that a useful answer needs, then runs two search sub-agents with deliberately different search strategies to build the map. It keeps following citation chains through OpenAlex and article links, stores originals under sources/, and marks each as paper, first-hand account, observation, digest or official document so you can decide what to trust. It does not summarize; a report script cuts quotations per claim and verifies that each excerpt matches the cited lines of the saved file. The goal is to avoid premature closure, which the README explains via library-science research on exploratory search.

The trade-off is real: you still have to read the sources yourself, and the output is a research folder under ~/research/<topic>/, not a ready answer. The stopping rule is also partly a heuristic — it halts when every claim has a source or when 70% of newly found material is already on the map, a threshold the README calls provisional. The author grounds each rule in cited papers, but the quality of a particular exploration is not something this input establishes.

Who should try it

Good fit: developers or writers who use Claude Code and need to reach primary material on a technical or academic topic — for example, following Gaussian Splatting papers and explainers, or tracing how experienced people sequence vocal mixing. A second mode collects new posts from tags and channels you found, which suits someone who wants a standing watch on a niche.

Poor fit: anyone who wants a synthesized answer, a one-shot brief, or a tool outside Claude Code. The skill is installed via the plugin marketplace or the skills CLI, with a manual copy as a fallback; other coding assistants are not covered. Those who need quick recommendations rather than sources will find the no-summary design counterproductive.

Requirements and limitations

Python 3.10+ with only the standard library runs the scripts. Optional tools change what is possible: pdftotext for PDF text, pdftoppm plus tesseract with Japanese language data for scanned PDFs, yt-dlp for YouTube subtitles, and ax for cleaner HTML-to-Markdown. Without them, the relevant source fails rather than producing a partial result.

The bigger ongoing considerations are privacy and network behavior. Queries are sent externally as search terms, so the README warns against confidential or personal topics. The fetch script refuses non-http(s) URLs and local or private addresses; if a VPN or proxy makes imozuru fail to fetch anything, the README documents IMOZURU_ALLOW_PRIVATE=1 to disable that check. The plugin carries no version number, so every commit to main counts as a new version; update via the plugin commands or enable automatic updates. Saved originals are for local reading and should not be redistributed.

HotRepo's take

The repository is 11 days old, and HotRepo recorded a #23 position on the NEW-REPOSITORY board (新项目榜, Python) on 2026-10-25. That is a dated snapshot, not a growth measurement, and comparable daily history is insufficient to judge whether interest persists. Once installed, check the README's saturation threshold and optional-tool table against the material you plan to research.