szmcp

Hello all, I wanted to share a little project I vibe-coded for myself that you may find useful. As many people here like to suggest, I wanted to give my small local LLMs access to information to improve their world knowledge. I didn't want to give my LLM free reign searching and browsing the web to keep my queries private and functional offline, so I wanted to give them an offline knowledge base. Wikipedia ZIM files from Kiwix were a good starting place for this. Several MCP servers for ZIM files exist, but I didn't like the existing ones I found for various reasons. The most notable one is https://github.com/cameronrye/openzim-mcp , which works in its advanced tool mode but has overly complicated context-bloating tools, and whose simple single tool mode doesn't work very well in practice. I built my own MCP server for ZIM files in Rust, exposing a simple tool set that's actually easy for small local LLMs to use, while providing all the functionality one normally needs. It's designed mainly for Kiwix MediaWiki ZIM archives generated by mwoffliner (such as Wikipedia, WIkivoyage, etc.) but also usable with many non-wiki ZIM files. I also wrote my own custom HTML to Markdown converter for MediaWiki pages that produces clean, well-formatted Markdown including special content such as wiki infoboxes, LaTeX formulas, tables, etc. It also strips out references and boilerplate sections from wiki pages to keep the resulting markdown clean and context efficient. You can hook this MCP server to llama.cpp's Web UI to give your small local LLMs much better world knowledge. A system prompt that I found works well is: You are a helpful assistant. When answering factual queries, search through Wikipedia using the provided ZIM access to ground your answers. If the articles or sections you read don't have relevant details, you can search more, but don't keep searching forever; you need to answer reasonably quickly. I tested it with various LLMs of varying sizes. I got good results with Gemma 4 12B (or bigger), IBM Granite 4.2 8B (or bigger), and Ling 3.0 Flash (best results while still maintaining usable speed on my 128 GB Mac). Qwen 3.6 35B-A3B was usable but tended to overthink and hallucinate; Qwen 3.8 27B was too slow to be usable for this purpose on my Mac. I also experimented with smaller models, and got usable results for simpler queries with MiniCPM5 2B, LFM 2.5 2.6B, and IBM Granite 4.2 3B. Gemma 4 E4B did not work well for this. I also build a sub-command within this tool to convert entire ZIM files from HTML to Markdown to save disk space (and avoid the need to convert on every tool call). It converts a 49 GB Kiwix nopic full English Wikipedia ZIM file into a 19 GB Markdown ZIM file, while maintaining all article content (aside from references) and maintaining full-text search. Likewise, it converts the 17 GB top-1M nopic enwiki Kiwix ZIM file from 19 GB to 6 GB. You can make the resulting ZIM files even smaller if you specify the option to only index article intros for full-text search (since the full-text search Xapian index is a large fraction of the file size). The converter is multi-threaded and written fairly efficiently using Rust, so you can convert all the millions of articles in a full English Wikipedia Kiwix files in a few hours on a typical modern computer. GitHub link: https://github.com/sultanqasim/szmcp