server-pdf: read_pdf_bytes hides the byte payload in structuredContent only — text content is a bare size summary, so CLI clients can't read the PDF

[jeius]

Summary

@modelcontextprotocol/server-pdf v2.0.0's read_pdf_bytes returns the requested bytes only inside structuredContent. The content[0].text block contains just a size summary like 256 bytes at 0/641720 with no data.

Clients that render only content[] text blocks (e.g. terminal/CLI MCP clients like opencode) never see the bytes — they are silently dropped. The description also warns "The model should NOT call this tool directly", yet it is the only server-side way to get raw PDF content when the interactive viewer path is unavailable.

Reproduction

1. Run the server: npx @modelcontextprotocol/server-pdf --stdio <allowed-dir>
2. Send an initialize handshake, then:

   {"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"read_pdf_bytes","arguments":{"url":"/path/to/file.pdf","offset":0,"byteCount":256}}}

3. Raw server response (verified by talking to the stdio transport directly):

   {
     "content": [{"type": "text", "text": "256 bytes at 0/641720"}],
     "structuredContent": {
       "url": "/path/to/file.pdf",
       "bytes": "JVBERi0xLjQ...",
       "offset": 0,
       "byteCount": 256,
       "totalBytes": 641720,
       "hasMore": true
     }
   }

   The base64 bytes decode to the real %PDF-1.4 header, so the data is correct — it's just unreachable to text-only clients.

Expected behavior

The byte payload should be reachable through the standard content[] text block, e.g. include the JSON payload (or the base64 bytes) in the text content, not only in structuredContent. Any MCP client that renders content[] should be able to consume the data without requiring structured-content support.

Related issue: display_pdf + interact requires an interactive viewer

display_pdf returns a viewUUID and interact get_text waits for the viewer iframe to poll within 8 seconds ("Viewer never connected for viewUUID ... (no poll within 8s)"). In headless/CLI contexts no iframe mounts, so text extraction via this path always fails. There is no non-viewer server-side text-extraction tool, so headless clients have no working path to PDF text today.

Environment

• Package: @modelcontextprotocol/server-pdf 2.0.0 (installed via npx, run with --stdio under Docker isolation)
• Client: opencode CLI (terminal MCP client)
• The server runs fine: connection, init, and tool discovery all succeed; only content delivery is affected.