Revenue Administration MCP Server

Turkish tax legislation changes constantly and lives in scattered places: statutes, communiqués, circulars, private rulings. The answer to a question almost always exists in writing - but finding it requires knowing the language of the legislation. A taxpayer asks "can I appeal this penalty?"; the legislation says "penalty reduction under VUK 376". That gap in language is exactly what stands in the way.

General-purpose language models are unreliable here. They invent article numbers, misquote rates, and blur repealed provisions with current ones. In a legal and fiscal setting a confident but wrong answer is more dangerous than no answer, because it looks authoritative.

The goal was an assistant where every statement traces back to an official source, and which refuses to answer anything it cannot ground there. It had to be reliable enough for a certified accountant to use on real client files.

I built a Model Context Protocol (MCP) server that connects to the official Revenue Administration legislation site at query time, and a web platform on top of it.

1. It translates the question into legal terminology first. Users write in everyday language. Before searching, the system converts that into the term the legislation actually uses - "what happens if I take money out of my company" → "dividend distribution withholding". This step is why the search works at all.
2. It reads live, not from a downloaded copy. Three tools - legislation search, category listing and full document text - pull directly from the official site, with results ordered newest first, because in law recency is decisive.
3. It doesn't invent when it can't find. If a search returns nothing it tries five formulations in turn: the term, the statute name, the article number, the abbreviation, a synonym. It won't say "not found" before at least three attempts - and if it still can't, it says so and points to the official legislation portal rather than guessing.
4. It answers in two layers. First the technical version - article, communiqué number, rate, date. Then the same information in plain Turkish, without jargon. Plain-language equivalents for terms like withholding, exemption and assessment are written into the system, as are distinctions practitioners routinely confuse.
5. Not just a server - a product. On top of the MCP server I built user accounts, a streaming chat interface, file upload, subscription and quota management, usage and cost reporting, and an admin panel. It can be used through its own web interface or from any MCP-compatible client such as Claude Desktop.

The system is in regular use today by a certified accountant, working on real client files. One example: on a ₺49,000 penalty issued for seven unfiled rental income declarations, the system found the relevant article and showed its source; the objection built on that article brought the penalty down to ₺10,500. The decision was the tax office's - the system's job was to find the grounds and show where they came from.

The same structure works in any field where the answer sits in a written source: legislation, contracts, technical specifications, internal procedures.

Stack: Python MCP server (live retrieval via Playwright), Node.js/Express backend, React front end, SSE streaming, Prisma + PostgreSQL, Docker.