Embed the server in your own program

Start the server from your own Node.js program when you need options the sezzlee-pdf command does not expose: how many extracted documents stay in memory, how many reads may run at once, or an OCR binding built in code.

Install the packages

mkdir -p ~/sezzlee-embed && cd ~/sezzlee-embed && npm init -y >/dev/null
npm install --silent @sezzlee/pdf-mcp @modelcontextprotocol/server

@modelcontextprotocol/server provides the stdio transport the server is served over.

Write the program

cat > ~/sezzlee-embed/server.mjs <<'EOF'
import { serveStdio } from "@modelcontextprotocol/server/stdio";
import { createDocumentRoot, createPdfMcpServer } from "@sezzlee/pdf-mcp";

const root = await createDocumentRoot(process.argv[2]);
const server = createPdfMcpServer(root, {
  documentCacheSize: 4,
  maxConcurrentExtractions: 1,
});
serveStdio(() => server);
EOF

createDocumentRoot checks the folder the same way the command does and throws if it does not exist. createPdfMcpServer takes these options, all optional:

Option	Default	Meaning
`documentCacheSize`	`2`	Extracted documents kept in memory, 1 to 16. Each one holds the Markdown of every page.
`maxConcurrentExtractions`	`2`	Documents extracted at once before a new read fails with `resource_limit`.
`maxConcurrentListings`	`4`	`list_documents` calls running at once before a new one fails with `resource_limit`.
`ocr`	none	An OCR binding, the same object a `--ocr` module exports by default.

A documentCacheSize outside 1 to 16 throws invalid_argument when the server is created.

Call it

The program is a stdio MCP server like the command, so the Inspector can start it:

npx -y @modelcontextprotocol/inspector --cli node ~/sezzlee-embed/server.mjs ~/sezzlee-pdf \
  --method tools/call --tool-name describe_document --tool-arg filePath=annual-report.pdf \
  | jq -c '.content[0].text | fromjson | {pageCount, documentType}'

Output

{"pageCount":4,"documentType":"text_based"}

In a client configuration, the command is node and the arguments are the program's absolute path and the folder.

Bind OCR in code

The ocr option takes the binding object directly, so the adapters from Read scanned pages with OCR can be configured without a separate module:

import { createPdfjsRasterizer } from "@sezzlee/pdf-raster-pdfjs";
import { createOllamaOcrProvider } from "@sezzlee/ocr-ollama";

const server = createPdfMcpServer(root, {
  ocr: {
    rasterizer: createPdfjsRasterizer(),
    provider: createOllamaOcrProvider({ baseUrl: "http://127.0.0.1:11434", model: "deepseek-ocr:3b" }),
    dpi: 200,
    maxPagesPerCall: 5,
    timeoutMs: 60_000,
  },
});

maxPagesPerCall and timeoutMs can only lower the server's own ceilings of 10 pages and two minutes. To use a different model host, implement the two methods yourself: render({ bytes, pages, dpi, signal }) returns { page, image, mediaType } for each requested page, and recognize({ pages, signal }) returns { page, markdown, confidence? } for each image. Page numbers are 1-based, and an answer for a page that was not asked for is ignored.