Read scanned PDFs and tables from an agent

A scanned PDF may contain images without useful embedded text. Docspeed provides distinct MCP tools for text geometry (ocr), ordered content and Markdown (parse), and table structure with optional cell grounding (tables_extract). Choose the tool according to the output you need.

Connect

Set DOCSPEED_API_KEY privately in the terminal used to launch your client. Add Docspeed to Codex:

codex mcp add docspeed --url https://api.docspeed.ai/mcp/ --bearer-token-env-var DOCSPEED_API_KEY

Or Claude Code:

claude mcp add --transport http --scope project docspeed https://api.docspeed.ai/mcp/ --header 'Authorization: Bearer ${DOCSPEED_API_KEY}'

Start a new session and approve the connection. The complete setup guide includes the companion skill and timeouts. Normal processing charges apply and documents leave your computer.

Choose the output

| Task | Tool | Evidence to preserve | | --- | --- | --- | | Read text and its position | ocr | Returned text and geometry | | Turn a document into ordered content | parse | chunks[].region_ids, regions, and partial-page failures | | Recover rows, columns, and cells | tables_extract with grounding: "cell" | Table/cell structure and returned grounding regions | | Extract named business fields | schema_auto, then extract | Field values, missing values, and their source regions |

Example prompt: “Read this scanned PDF and extract every table. Preserve merged headers and supporting cell regions. Report unreadable pages explicitly.”

Use the direct streaming upload helper and reuse the returned file_id. Reserve files_upload for tiny inputs at most 4 KiB. The hosted server cannot open your local path. Parse exposes HTML for tables when available; do not flatten merged cells into pipe Markdown and discard structure. A failed OCR page must remain visible; the server does not silently switch to a classic OCR mode unless that mode was explicitly selected.

Inspect evidence before adopting

The recorded synthetic invoice example contains line-item cells and source overlays from Extract. It demonstrates the evidence review experience; it is not a scanned-document OCR benchmark or a Tables recording. No speed or accuracy comparison is claimed here.

Supported uploads are PDF, PNG, JPEG/JPG, TIFF/TIF, WebP, BMP, and GIF, including multi-page TIFF. Native Word and Excel workflows require beta API access by contacting Docspeed. See OCR and Layout, Tables, invoice extraction, privacy, and support.