A PDF reader with two ways to read with LLMs:
- Threads — select a passage, start a discussion in the margin.
Invite different models with
@mentions(@claude,@grok,@gpt,@code). Inspired by Fermat's Library. - Chapter summaries — drop a PDF, get a standalone interactive site for the chapter — one section per concept, interactive figures, 3D models, math, review flashcards. Live demo →
(4× speed — threads)
Everything stays in your browser — PDFs, threads, settings. No server-side database.
- Multiple models in one conversation —
@mentiondifferent agents to bring them into the thread. - Built-in tools — the agents can search inside the paper, look up arXiv references, browse web pages, run Python, scroll the viewer, and highlight passages for you.
@code— lets Claude run shell commands and edit files in a dedicated workspace, for deeper paper-aware tasks.
A planner reads the chapter, decides the sections, then runs one writer per section in parallel. Each writer produces a self-contained interactive page with figures, math, 3D models, definitions, and review cards. The output is a static website — that's literally what the live demo is.
Roughly per ~40-page chapter:
| Model | Cost |
|---|---|
| Claude Opus | ~$1.50 |
| Kimi K2.6 (OpenRouter) | ~$0.30 |
| GLM-5.1 (OpenRouter) | ~$0.50 |
| DeepSeek V4 (OpenRouter) | ~$0.20 |
In ⚙ Settings, pick a planner + writer model, drop a PDF, click Summarize.
curl -fsSL https://raw.githubusercontent.com/eitanporat/paperchat/main/install.sh | bashThe installer clones to ~/paperchat, installs dependencies, and asks
for an OpenRouter API key. After that:
- Open http://localhost:5173.
- Drop a PDF.
- For a thread: select a passage and click
@claude,@grok, or@gpt. On macOS,@codeworks automatically against your localclaudeCLI. - For a chapter summary: open ⚙ Settings, pick a planner + writer
model, then click Summarize on the chapter chip. Output lands in
cc-workdir/chapter-<id>/.
Manual install:
git clone https://github.com/eitanporat/paperchat.git
cd paperchat
npm install
npm run devThe dev server is intended for localhost only — don't expose it on
a public interface. @code runs with full filesystem and shell access
inside its sandbox directory. Treat as semi-trusted.
MIT — see LICENSE.