server-pdf: get_viewer_state always returns pageCount: 1 and get_text returns empty in VS Code

[EliottDelhaye]

Summary

When using @modelcontextprotocol/server-pdf (v1.6.0) in VS Code with GitHub Copilot Chat, display_pdf opens a viewer but get_viewer_state always reports pageCount: 1 and get_text returns empty content for all PDFs -- local files and remote URLs alike.

---

Expected Behavior

• get_viewer_state should return the correct page count
• get_text should extract text content from PDF pages
• get_screenshot should return a page image

---

Actual Behavior

For every PDF tested:

• get_viewer_state returns { currentPage: 1, pageCount: 1, zoom: 100, ... }
• get_text returns --- Page 1 --- with no text content
• get_screenshot returns ERROR: No screenshot returned
• The totalBytes in the display_pdf response is correct, suggesting the file is loaded by the server

---

Reproduction Steps

1. Configure the server in VS Code user mcp.json:

   {
     "servers": {
       "pdf": {
         "command": "npx",
         "args": ["-y", "@modelcontextprotocol/server-pdf", "--stdio"]
       }
     }
   }

2. Open GitHub Copilot Chat (Agent mode)
3. Ask it to display any PDF, for example the default arxiv paper:

   display_pdf with url https://arxiv.org/pdf/1706.03762
   

4. Then ask it to get the viewer state or extract text:

   get_viewer_state
   get_text page 1
   

5. Observe: pageCount is 1, text is empty

---

PDFs Tested (all show same behavior)

PDF	Type	Pages (pdfinfo)	Pages (MCP)	Text (MCP)
Local LaTeX Beamer slides (362x272 pts)	Local file	51	1	empty
Local pdflatex A4 article	Local file	3	1	empty
https://arxiv.org/pdf/1706.03762 (Attention paper)	Remote URL	15	1	empty

All PDFs have verified text content via pdftotext and correct page counts via pdfinfo.

---

Environment

• VS Code: latest stable (April 2026)
• GitHub Copilot Chat extension: v0.44.1
• server-pdf version: 1.6.0 (latest)
• OS: macOS (Apple Silicon)
• Node: via npx

---

Analysis

The display_pdf tool returns a correct totalBytes value, so the server reads the file. The issue appears to be that get_viewer_state and get_text query the PDF.js viewer (MCP App), but the viewer either hasn't finished loading the document or isn't properly communicating its state back through the command queue.

This could be a timing issue (commands sent before PDF.js finishes rendering) or a VS Code MCP App host issue where the iframe doesn't properly initialize.