tax-law-mcp

What it does

tax-law-mcp gives an assistant the original wording of Japanese tax law so that answers about articles and circulars can be checked against the text instead of recalled from memory. It has seven tools in three groups. Statutes come from the Digital Agency's e-Gov law API v2 and are returned as Markdown, down to paragraph, item and sub-item, with supplementary provisions too; 26 major laws (Income Tax Act, Corporation Tax Act, Consumption Tax Act, the Special Taxation Measures Act and their orders, plus the Civil Code and Insurance Act) are preset and can be named by common abbreviations such as 所法 or 措法, and any other law can be found by keyword or given by law_id. Interpretive circulars (通達) are scraped from the National Tax Agency site, 17 of them, with a table-of-contents tool. Published rulings of the National Tax Tribunal (kfs.go.jp), about 1,950 cases across 97 volumes, can be listed by tax type, searched by keyword and read in full.

It is written in Japanese: tool descriptions, inputs and output are Japanese. The server also sends the client a long set of instructions telling the assistant to quote fetched text verbatim with its source URL, to run up to four rounds of checking, and to mark anything found through web search as secondary information. Those instructions state that court judgments (判例) are outside what it covers (src/server.ts). The same repository also deploys to Vercel as a remote MCP endpoint and a REST API with an OpenAPI file, but there is no public hosted instance: you deploy your own.

What it needs

Node.js with npx (the package's dependencies include @modelcontextprotocol/sdk; the repository's dev setup targets Node 20). No account or API key: it calls public government sites (laws.e-gov.go.jp, nta.go.jp, kfs.go.jp), so outbound internet access is needed. The remote and REST mode needs your own Vercel deployment (vercel deploy).

Limits

Covers Japanese national tax law only, and only the 17 listed NTA circulars; court judgments are not included. search_saiketsu returns 10 results by default, at most 30 with limit, and matches only the case summaries and category names; its first full search fetches all 97 volume indexes and takes 30-60 seconds (narrow it with latest), after which results are cached in memory for an hour. Circulars and rulings are scraped from web pages, so a change to those sites can break them. The README warns that Claude's remote (custom connector) MCP connection has a known fault and recommends the local npx setup instead.

Who it suits

Suits tax accountants, lawyers and developers working in Japanese who want an assistant to cite the actual article, circular or tribunal ruling rather than paraphrase from memory. It is read-only, needs no key and runs locally, so it is easy to try. It is not for other countries' tax law, nor for case law from the courts, and English-only users will find its output hard to use. For Korean tax law we hold korea-tax-law-mcp, a separate server with a hosted endpoint.