8 个 GitHub 仓库拼出 Claude + Jev 研究台
用 Crawl4AI、Docling、Zotero、Qdrant、jev-mcp、LangGraph、Langfuse、Streamlit 搭研究流程：Jev 先筛候选材料，Claude 再深挖并写简报，最后人工核对来源。
A RESEARCH DESK IN 8 GITHUB REPOS. READ THE WEB, OPEN PDFS, KEEP THE SOURCES. Competitor briefs. Product research. Technical deep dives. That folder of PDFs you keep meaning to read. Here's the stack I'd put around Claude + Jev: 01 Crawl4AI → turn web pages into readable Markdown, with links preserved. https://github.com/unclecode/crawl4ai 02 Docling → pull structured content out of PDFs, including layouts, tables and scanned pages with OCR. https://github.com/docling-project/docling 03 Zotero → keep the source library organized so you can find and cite things again. https://github.com/zotero/zotero 04 Qdrant → search stored material by meaning, with metadata attached to each result. https://github.com/qdrant/qdrant 05 jev-mcp → give Claude Jev tools for classification, scoring and checks. Its triage tool can screen files before the agent opens them. https://github.com/rashedInt32/jev-mcp 06 LangGraph → connect the steps, keep workflow state and pause for human review. https://github.com/langchain-ai/langgraph 07 Langfuse → inspect the calls and traces when the agent does something weird. https://github.com/langfuse/langfuse 08 Streamlit → build a screen where you can inspect the material and review the draft. https://github.com/streamlit/streamlit The workflow I'd start with: collect the material → let Jev sort the candidates → have Claude investigate the shortlist and write a brief → open the sources yourself before approving it. You'd still need to wire the pieces together and budget for model APIs or hosting. For a small first version, start with one question and five sources. Add the bigger search library when you actually need it. Save the repos. The Jev article below explains the decision layer I'd use between collecting everything and asking Claude to read it.
这条是引用着另一条写的：x.com/e1genqq/status/2110016031917855006