imozuru: research an unfamiliar field down to the original sources
Added Updated NewCopyable prompt below
Published Last verified Released
Our verdict
Use it when a quick A.I. summary is not enough and you need to check what a claim rests on. Instead of reading the top few search results, imozuru first learns the field's terms, key people and venues, then follows references, citing papers, authors and blog links until every point of your question has a source. It saves the sources and tells you which file and which lines support each point, and a script refuses to build the extract page if any quote isn't word for word. Builders use it to fact-check viral claims, find papers they had missed in a neighbouring field, and dig through interview articles; one ported it to an MCP server for Hermes Agent. Two costs: it is slow and token-hungry (one builder measured 360k tokens and 22 minutes against 178k and 6 minutes for a plain request, with 9 of 9 spot-checked citations holding up), and its instructions, script messages and dashboard labels are in Japanese; the README says it has only been tried with requests written in Japanese.
- Who it's for
- Founders, analysts, writers and builders using Claude Code who need to get up to speed on an unfamiliar topic and want every claim traceable to a source
- Leverage
- A saved, line-referenced source pack and a layered explainer for a new field, instead of a summary you cannot check
- Effort
- Medium
- Date added
What imozuru does
imozuru (MaryCache/imozuru, MIT) is a Claude Code skill and plugin, also usable in Codex CLI, based on information-seeking research from library science (Bates's berrypicking, Ellis, Kuhlthau). It writes the question and 3 to 6 points needed to answer it, sends two web-search subagents with different approaches to map terms, people and venues into map.md, then follows citations (via the free OpenAlex API, no key), links and authors. Sources are saved to sources/ with a label, and 404s and login walls are dropped. It stops when every point has a source or when 70% of the new finds are already on the map. Version 2.0 (25 October 2026) added a detailed report at three reading levels, where a script checks every quoted sentence against the saved sources, plus a dashboard of all your investigations that you can open on a phone over Tailscale. A watch mode (/imozuru watch <topic>) collects new posts from the venues a dig found via RSS. Since 2.1.0, plugin users get updates only on numbered releases; 2.2.0 (28 October) added Codex CLI instructions and a pdfminer.six fallback for PDFs, and 2.2.1 (18 November) tells the agent to open outside pages only in a sandboxed browser. The scripts need only Python 3.10+; pdftotext, tesseract, yt-dlp and ax are used when installed.
How builders use imozuru
- Install it as a Claude Code plugin (/plugin marketplace add MaryCache/imozuru, then /plugin install imozuru@imozuru) or with npx skills add MaryCache/imozuru.
- In Codex CLI, link skills/imozuru from a clone into ~/.agents/skills/imozuru and call it with $imozuru.
- Ask it to dig into a topic you need to decide on (a technique, a market, a standard) and read the extract page before trusting the explainer.
- Open the sources and line numbers it points to for the claims that matter, rather than re-searching.
- Run /imozuru watch <topic> afterwards to follow new posts from the venues it found, and budget it for questions that deserve a deep dig.
What a dig leaves on your disk
From the English README, checked 19 Nov 2026. Everything lands in one folder per topic, for example ~/research/<topic>/.
- map.md: the field's terms, people and venues, plus a table of saved sources with their original URLs.
- sources/: the saved originals, each labelled so you can judge how far to trust it.
- The extract page: passages cut from the sources, point by point. report.py checks that each one appears verbatim at the stated lines and won't build the page if a single word differs.
- The detailed report: an explanation of the field at three reading levels, every sentence backed by a quote that digest.py has checked, with the full record of what was searched.
- The dashboard: every investigation, why it stopped and where its sources came from. It can be served on 127.0.0.1 behind a passphrase and opened from your phone through Tailscale.
How it labels sources
Each saved source gets one label; which to trust is left to you. From the README, checked 19 Nov 2026.
| Label | What it means |
|---|---|
| Paper | Peer-reviewed or conference papers and arXiv preprints |
| First-hand | A record by the person who did the work: trying, making or measuring it |
| Observation | Notes on someone else's work or methods |
| Roundup | Articles that collect or restate others' findings |
| Official | Documents from the maker of the tool or product |
Imozuru: limits and gotchas
- Your question leaves your machine as search terms, so the README says not to use it for confidential or personal data.
- No paywall workarounds: sources it can't reach are reported as not reached.
- Japanese interface: SKILL.md, script messages and the extract page and dashboard labels are in Japanese, and the skill has only been tried with Japanese requests.
- The 70% saturation threshold for stopping is provisional, set through trial runs.
- PDFs need pdftotext or pdfminer.six (slower, and large files can hit a 300-second limit); scanned Japanese PDFs also need tesseract's Japanese data.
- It never fetches local or internal addresses; behind some VPNs and proxies that blocks everything until you set IMOZURU_ALLOW_PRIVATE=1.
- Saved sources are local reading copies; the README asks you not to redistribute them.
Imozuru compared with other tools here
- GPT Researcher: plans questions, searches and writes a cited report with the model you choose. Faster; imozuru goes deeper into original sources and checks quotes.
- Open Deep Research: LangChain's deep research agent with parallel sub-agents. A better base if you are building your own research pipeline.
- Gemini Deep Research: a hosted report from Google with nothing to install. No line-level citations to files you keep.
- last30days: what people said about a topic in the last 30 days. For current sentiment, not a field's original sources.
- Perplexity API: live web answers with citations from an API, for apps and quick lookups.
When to pick imozuru
- Pick imozuru if you are new to a field and need every claim traced to a file and line you can open.
- Pick GPT Researcher or Open Deep Research for a cited report in minutes with the model and search you choose.
- Pick last30days for what people are saying about a topic right now.
- Skip it for quick lookups or confidential topics.
Copy this: Install + first dig
# Official source: github.com/MaryCache/imozuru (MIT). Needs Python 3.10+.
# In Claude Code:
/plugin marketplace add MaryCache/imozuru
/plugin install imozuru@imozuru
# or: npx skills add MaryCache/imozuru
# Optional helpers (used when installed): poppler (pdftotext), tesseract, yt-dlp
# Ask (the README's own example, in English):
Dig into this: I keep hearing about Gaussian Splatting but have no idea what it is. Dig through the papers and explainers.
# Results: ~/research/<topic>/ (map.md, sources/, report.html, digest.html)
# Dashboard of every investigation:
python3 skills/imozuru/scripts/dashboard.py ~/research
# Follow new posts from the venues it found:
/imozuru watch <topic>Imozuru questions
Does imozuru work in English?
Claude reads its instructions in any language and the README has an English version, but the skill's messages and page labels are Japanese and it has only been tried with Japanese requests.
Does it need API keys?
No. Citation following uses the free OpenAlex API without a key, and web search runs through Claude Code or Codex. The scripts need Python 3.10+.
How does it decide when to stop?
When every point of your question has a location in a source, or when 70% or more of what the two searchers find is already on the map.
Can it keep following a topic?
Yes. /imozuru watch <topic> collects new posts via RSS from the venues a dig found, such as blogs, YouTube and arXiv.
Related discoveries
See the newest picks →- ToolResearch & Automation
last30days: research any topic from the last 30 days
A free, open-source agent skill (MIT) for Claude Code, Codex, Cursor, Gemini CLI and other Agent Skills hosts: give it a topic and it searches what people posted in the last 30 days on Reddit, Hacker News, X, YouTube, Polymarket, GitHub and more, then writes one brief with links to the sources.
NewEffort: LowPrompt#Plugin / extension#Open source
- ToolResearch & Automation
Open Deep Research (LangChain)
LangChain's MIT deep research agent on LangGraph: plans, researches with parallel sub-agents and writes a report, with the models and search you choose.
Effort: High#Open source
Get the best new A.I. picks in your inbox
The best new agent workflows and tools we’ve vetted, with a clear verdict on each. Short, occasional, useful.