rosetta

MCP server for searching MikroTik RouterOS documentation. Gives your AI assistant searchable access to 317 documentation pages, 4,860 property definitions, 40,000-entry command tree, and 144 hardware product specs — with direct links to help.mikrotik.com.

Tested with Claude Desktop, Claude Code, VS Code Copilot (including Copilot CLI), Cursor, and OpenAI Codex on macOS, Linux, and Windows.

Most retrieval-augmented generation (RAG) systems use vector embeddings to search documentation. This project takes a different approach: SQLite [FTS5](https://www.sqlite.org/fts5.html) full-text search as the retrieval layer — what we call SQL-as-RAG.

For structured technical documentation like RouterOS, full-text search with BM25 ranking beats vector similarity. Technical terms like "dhcp-snooping" or "/ip/firewall/filter" are exact tokens — porter stemming and proximity matching handle the rest. No embedding pipeline, no vector database, no API keys. Just a single SQLite file that searches in milliseconds.

The data flows: HTML docs → SQLite extraction → FTS5 indexes → MCP tools → your AI assistant. The database is built once from MikroTik's official Confluence documentation export, then the MCP server exposes 11 search tools over stdio or HTTP transport.

• 317 documentation pages from MikroTik's official help site (~515K words)
• 4,860 property definitions with types, defaults, and descriptions
• 5,114 commands in the RouterOS command hierarchy (551 directories, 34K arguments)
• 1,034 callout blocks — warnings, notes, and tips with important caveats
• 144 hardware products — CPU, RAM, storage, ports, PoE, wireless, license level, pricing
• 46 RouterOS versions tracked (7.9 through 7.23beta2) for command history