go-docs-mcp 1.4: spreadsheets, slides and ebooks

More than PDF

go-docs-mcp started as a way to let an AI client read the PDF files in a folder. Version 1.4.0 opens it to the documents people keep next to their PDFs. With pandoc installed, it reads Word (DOCX), Excel (XLSX), PowerPoint (PPTX), OpenDocument text (ODT), EPUB, HTML, RTF and Jupyter notebooks (IPYNB), and convert_to_markdown turns any of them into GitHub-flavoured Markdown.

TXT, Markdown and CSV need no extra program. Images in PNG, JPG, WebP, TIFF and BMP go through OCR, and list_documents now lists them too, each entry with its type.

The server still has 13 tools, all read-only. Every call that worked with v1.3.0 still works, because the new parameters are optional.

Long documents, read by pages

A 400-page manual does not fit in one answer. read_document, convert_to_markdown and ocr_document take offset and max_chars (50,000 characters by default), and each page of text ends with the offset to ask for next. search_document takes max_results and extract_images takes max_images.

The server also tells the model the order that works best: list the documents, read the outline, search, and only then read by pages.

Images you can see

extract_images used to return raw image data as text, which no client could show. It now returns real PNG images, scaled to 1600 px on the longer side. They are also faster to make: about 0.5 s instead of 2.3 s for a 2048 x 2048 photo.

Documents in Claude Code

Every document in the folder is a docs:// resource, so in Claude Code you can type @ and pick a file by name. The tools that return lists, metadata or tables also give structured output, for clients that read fields instead of text.

Safer by default (v1.3.0)

v1.3.0 came out earlier the same day. It did not add or change any tool; it closed gaps:

• Documents stay inside the folder you give the server. A symlink that points outside it is no longer listed or read.
• Every external program (poppler, tesseract, pandoc) runs with a time limit and is stopped when it goes over it.
• read_url has a 60 s timeout, refuses private and cloud metadata addresses, and can be removed from the server with DOCS_MCP_DISABLE_URL=1.
• Error messages no longer show paths of your machine.

Install

On macOS and Linux, Homebrew installs poppler, tesseract and pandoc with it:

brew install drolosoft/tap/go-docs-mcp

For Claude Code there is also a plugin in the Drolosoft marketplace. It needs the binary installed first:

claude plugin marketplace add drolosoft/claude-plugins
claude plugin install go-docs-mcp@drolosoft-marketplace

Archives for macOS, Linux and Windows, with a checksums.txt file, are on the v1.4.0 release page, and go install github.com/drolosoft/go-docs-mcp@latest still works. Every change is listed in the changelog.