JoeZhang-0x000/wenqu

What it does

Wenqu is a local, agent-native assistant for reading papers. It keeps a personal paper library, answers questions about selected passages with page citations, and converts papers into two-host podcasts with editable scripts and synchronized transcripts. It suits researchers and engineers who accumulate PDFs but have limited desk time and want to listen on a commute or during chores.

The work is split across three areas: library management through an approval-card flow for adding and tagging papers, in-page asking with answers that cite page numbers and can be checked against a paper's code, and audio generation from a script. Agent interaction is handled by an external CLI backend, while speech synthesis uses Doubao's API.

Quick start

Prepare Node.js 20 or later, one supported agent CLI (Claude Code, Codex, pi, or DeepSeek Harness), and a Doubao Speech API key if you want audio.

git clone https://github.com/JoeZhang-0x000/wenqu.git
cd wenqu
npm install
npm run build
npm start

The browser opens automatically. In Settings → Agent backend, choose a backend and click Test; then enter the Doubao key under Settings → Voice API and test the connection. Add papers as PDFs or arXiv links, or drag a PDF onto the page. Run npm link if you want the wenqu command available from any directory. Setup details, collections, tags, and file locations are covered in the guide.

Usage constraints

The project is early-stage and has been developed and tested on macOS only. Scanned PDFs without a text layer are not supported. Code reading is limited to GitHub repositories. Podcast voices come from Doubao and are Chinese by default; English and Japanese synthesis is also possible. AI-generated content can be wrong and should be checked against the source before publication.

Paper content is sent to the chosen agent service and to Doubao only when audio is synthesized. The server listens on 127.0.0.1, and agents run in an empty temporary directory with restricted permissions. Changes made by the assistant require explicit approval. Papers, chats, and episodes are stored as plain files in a vault, with the API key stored outside it. The full security model and limitation list are in the guide.

HotRepo's take

Wenqu entered HotRepo's new-project list on 2026-11-11 with 103 stars; the recorded growth window was only about 90 minutes long and added 3 stars, so the listing shows early attention rather than a sustained trend. The repository has no tracked GitHub daily rank. The interesting technical bet is the separation of duties: a general agent CLI handles reading and asking, while the app enforces local-only access, sandboxed agent runs, and human approval for every change. That makes it a useful early adopter choice for paper workflows on macOS, with cross-platform support and scanned-PDF handling still undefined.