Read and mark up a PDF together with the AI.
You and the AI work on the same PDF in a viewer inside the chat: it searches, highlights, annotates, stamps and fills in forms while you watch, then saves the finished copy.
What it is
An example server from the MCP Apps project, published as @modelcontextprotocol/server-pdf. MCP Apps let a server show an interactive interface inside the chat; this one shows a PDF viewer. It opens local PDFs and papers from academic sources such as arXiv, bioRxiv and Zenodo, and keeps the AI informed about the page you are looking at.
What you get
- Available local files and allowed online sources listed
- A PDF opened in an interactive viewer, inline or full screen
- Navigation, text search, and the text or a screenshot of chosen pages
- Highlights, notes and stamps added to pages
- Form fields filled in
- The annotated PDF saved to a new file, or over the original local file
Requirements
Node.js 20 or newer; no account or key. To see the viewer and use its interactive actions, an MCP client that displays MCP Apps, such as Claude Desktop. It reads only local files under the folders your client shares with it. Saving over Claude Desktop attachments is refused unless the server is started with --writeable-uploads-root.
Setup effort
One command — npx -y @modelcontextprotocol/server-pdf --stdio