root-ext-docs
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@root-ext-docs extract text from the meeting notes PDF"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
root-ext-docs
Document-engineering MCP tool server for ROOT Workspace (and any MCP host). Twelve tools across four surfaces:
PDF —
pdf_info(which pages are text, which are scans),pdf_extract_text(page-ranged),pdf_pages_to_images(render toderived/— where the workspace viewer and the model's image-read path pick them up),ocr_pdf(local tesseract).Office —
docx_extract,pptx_extract: structure-keeping markdown (headings, tables, slide notes).Images —
ocr_image,image_convert,image_crop,image_annotate(labelled rectangles for pointing at drawings).Archives —
archive_list,archive_extract(zip-slip refused by name, size/entry caps).
Charter, binding and tested: extract and derive, never mutate
sources. No tool overwrites, deletes, or edits a user's file in
place; write tools refuse output paths that collide with inputs and
default to derived/ beside the source.
OCR needs the system tesseract binary (brew install tesseract) and
refuses by name without it. Nothing in this pack reaches the network.
Develop
uv venv && uv pip install -e ".[dev]"
.venv/bin/pytestRelated MCP server: pdf-reader-mcp
Package (signed .rootx, ADR-0018)
uv build
.venv/bin/python tools/build_pack.py # --unsigned for a test buildAvailable Tools
12 toolsarchive_extractAIdempotent
Extract an archive (or just members) into out_dir — default
derived/<name>/ beside it, and NEVER anywhere else: absolute
paths, .. members, and links are refused by name (zip-slip),
entry count and total size are capped so a zip bomb is an error,
not an outage.
Inputs: path (required), members, out_dir.
archive_listARead-only
List a zip/tar archive's entries (name, size, kind) plus the total uncompressed size — look before extracting. Bounded at 1000 entries; truncation is marked.
Inputs: path (required).
docx_extractARead-only
Extract a DOCX in document order with structure kept: headings as markdown levels, paragraphs, and tables as markdown tables — so you can cite the table under a section instead of fishing in text soup. Bounded; truncation is marked.
Inputs: path (required).
image_annotateAIdempotent
Draw labelled rectangles on a copy of an image (up to 50 marks, each {x, y, w, h, label?}) — point at regions of a drawing or screenshot when words alone are ambiguous. Writes a NEW file; refuses an out_path that collides with the input (charter).
Inputs: path (required), marks (required), out_path (required).
image_convertAIdempotent
Convert/resize/rotate an image into a NEW file — format by out_path suffix (png/jpg/webp/gif/bmp/tiff), optional max_dim bound on the long edge, rotation in 90° steps. Refuses an out_path that collides with the input (charter).
Inputs: path (required), out_path (required), max_dim, quality, +1 more.
image_cropAIdempotent
Crop a pixel box (x, y, width, height from the top-left) into a NEW file — e.g. isolate one view of a drawing before OCR or a close look. Refuses a box that leaves the image and an out_path that collides with the input (charter).
Inputs: x (required), y (required), path (required), width (required), +2 more.
ocr_imageARead-only
OCR one image (png/jpg/gif/webp/bmp/tiff) with the LOCAL tesseract binary — needs tesseract installed and refuses by name without it; nothing reaches the network. OCR misreads numbers — verify critical values by looking at the image itself.
Inputs: path (required), lang.
ocr_pdfARead-only
OCR scanned PDF pages (up to 10 per call): renders each page and
runs the LOCAL tesseract binary — needs tesseract installed
(brew install tesseract) and refuses by name without it; nothing
reaches the network. Use pdf_info first to find which pages need
OCR. OCR misreads numbers — verify critical values against the
rendered page.
Inputs: path (required), dpi, lang, pages.
pdf_extract_textARead-only
Extract the text layer of a page range (1-based: "3", "1-5", "2,4,9-11"; up to 50 pages per call). Page-ranged so a 400-page standard is read in slices instead of one blind gulp. Returns nothing useful for scanned pages — pdf_info names those, ocr_pdf reads them. Bounded; truncation is marked.
Inputs: path (required), pages.
pdf_infoARead-only
Inventory a PDF before reading it: page count, per-page size (mm), and — the part that matters — which pages carry a text layer and which are scans/drawings. Use it to decide between pdf_extract_text (text layer), ocr_pdf (scans), and pdf_pages_to_images (look at drawings). Reads the file on this machine; nothing leaves it.
Inputs: path (required).
pdf_pages_to_imagesAIdempotent
Render PDF pages to PNG/JPEG files (up to 20 pages per call,
dpi 30–400) in out_dir — default derived/ beside the source,
never on top of it (charter). The rendered pages are ordinary
images: open them in the workspace, or read them with fs.read to
LOOK at a drawing-heavy page the text extractor can't serve.
Inputs: path (required), dpi, pages, out_dir, +1 more.
pptx_extractARead-only
Extract a PPTX slide by slide: title, body text, tables, and speaker notes (quoted, marked as notes). Bounded; truncation is marked.
Inputs: path (required).
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 12 tools
Each tool targets a distinct resource or operation: archives, PDF (with clear separation between info, text extraction, rendering, and OCR), DOCX, PPTX, and images. The descriptions explain precisely when to use each, minimizing overlap.
Most tools follow a resource_action pattern (e.g., pdf_extract_text, image_crop), but ocr_pdf and ocr_image invert this to action_resource. pdf_info uses a noun as a verb. The pattern is largely consistent with minor deviations.
At 12 tools, the set is well-scoped for a document extraction server. Each tool has a clear role, covering archives, multiple document formats, and image processing without unnecessary duplication.
The surface covers core extraction workflows: listing/extracting archives, reading PDFs (text, renders, OCR), extracting DOCX/PPTX, and image processing. Minor gaps exist (no archive creation, no PDF metadata extraction) but do not hinder primary use.
Related MCP Connectors
Document conversion MCP server: PDF to Markdown, image OCR, spreadsheet parsing.
Document-to-Markdown MCP server — convert PDF, Office and HTML into LLM-ready Markdown.
Document processing over MCP: merge, split and compress PDFs, run OCR, extract document text.
Open-source developer, PDF, image, SEO and security tools over a stateless HTTP MCP server. No key.
Related MCP Servers
- AlicenseBqualityDmaintenanceA universal MCP server for document processing, conversion, and automation. Handle PDF, DOCX, HTML, Markdown, and more through a unified API and toolset.1328 npm140MIT
- FlicenseNot gradedqualityDmaintenanceMCP server for extracting text from PDF files, supporting local files and URLs.-
- AlicenseNot gradedqualityDmaintenanceUniversal MCP server for extracting text from various document formats including PDF, Excel, Word, CSV, and more, with support for streaming, limits, and markdown conversion.3MIT
- AlicenseNot gradedqualityCmaintenanceUnified MCP server that exposes GroupDocs.Total document-processing tools (annotation, comparison, conversion, merging, metadata, parsing, redaction, signing, watermarking, and more) as AI-callable tools for MCP agents.MIT