Featured image of post Inside korean-law-mcp — folding 41 government APIs into 17 MCP tools
Devtools

Inside korean-law-mcp — folding 41 government APIs into 17 MCP tools

A look at how korean-law-mcp wraps the entire Korean legal system in one MCP server and the tool-compression and anti-hallucination citation design behind it

Overview

korean-law-mcp is a TypeScript server that compresses 41 Open APIs from Korea’s Ministry of Government Legislation into 17 MCP tools. It is not a thin API wrapper — it ships citation verification that catches statute numbers an LLM invented, an impact-graph analysis that traces the ripple effect of a single article, and an automatic diff between two points in legislative time. Built by a civil servant who got tired of “searching the legislation portal manually for the hundredth time,” the project is an instructive case study of one MCP design principle: tool count is not feature count.

Why the project had to exist

South Korea has over 1,600 statutes in force, more than 10,000 administrative rules, and a sprawling body of rulings spanning the Supreme Court, the Constitutional Court, the Tax Tribunal, and the Korea Customs Service. All of it lives on one portal, but for a developer trying to reach the data programmatically the experience is rough. The legislation portal’s Open API exposes 41 endpoints behind a single free auth key (OC), but handling 41 endpoints is one problem and turning them into tools an LLM can use is a completely different one.

The Model Context Protocol is an open standard Anthropic released in November 2024 — a “USB-C port” that connects AI applications to external data and tools. Claude Desktop, Cursor, Visual Studio Code, Zed, and many other clients can call MCP servers. korean-law-mcp layers an MCP server on top of the legislation APIs, turning a one-line natural-language question into a chain of calls across those 41 endpoints.

From 89 tools down to 17 — the compression idea

The project’s central design decision happened in the move from v2 to v3. v2 took the intuitive route — one tool per API — so 41 APIs fanned out into 89 tools. The problem is the LLM’s point of view. An MCP client loads every tool’s JSON Schema into context at session start, and 89 schemas weighed in around 110 KB, burning half the context window on the tool list alone.

v3’s reframe is a dispatch table plus domain enum pattern: collapse tools that share a shape behind a single domain parameter. Seventeen ruling domains — case law, the Constitutional Court, the Tax Tribunal, the Fair Trade Commission, labor commissions, and more — merged into just two tools, search_decisions(domain) and get_decision_text(domain). The count dropped from 89 to 15 (now 17), and context cost fell from roughly 110 KB to 20 KB, an 82% cut. The notable part: not a single existing handler function was modified — only a new dispatch layer went on top.

Government APIsv2v3
API / tool count418917
AI context cost-~110 KB~20 KB
Feature coverage-100%100%

The remaining specialist tools (legal terminology, annexes and forms, amendment histories) did not disappear. A discover_tools → execute_tool proxy pattern lets the LLM search and call them only when needed. That is the MCP tool-discovery pattern applied to the legal domain — keep the exposed surface small while holding feature coverage at 100%.

Anti-hallucination — verify_citations

verify_citations, added in v3.5, is the most domain-native feature in the project. Legal answers generated by ChatGPT or Claude easily mix in plausible-but-nonexistent statute numbers — the familiar hallucination problem. verify_citations extracts article citations from user text with a regex, looks back 30 characters to recover the statute name, then cross-checks against the official government database in parallel. Results sort into three buckets: ✓ (exists), ✗ (does not exist, with the valid range shown), and ⚠ (statute name ambiguous).

What is interesting is the bugs that surfaced while empirically validating the feature. The v3.5.3 release notes describe testing five citations against the live government API and finding three false negatives. A substring mismatch where “Civil Act” matched “Refugee Act”; a parsing failure where the API returns paragraph numbers as enclosed Unicode numerals like "① " and parseInt stripped them into NaN; a short statute name (“Commercial Act”) buried at result 34 and dropped. The tool built to catch hallucinations nearly hallucinated itself — and the README keeps the full debugging trail that pushed it back to 5-out-of-5 accuracy.

v3.5.4 went further, introducing machine-parseable markers like [NOT_FOUND] and [HALLUCINATION_DETECTED] on every failure response. The trigger was real-world feedback: “in practice the AI keeps not finding things and then just answers however it wants.” The root cause was that some tools did not set the isError flag on a failed lookup, so the LLM never detected the failure and generated an answer anyway. It is a sharp reminder that getting an MCP tool to clearly signal failure to an LLM is harder than it sounds.

v4.0’s three killer features

v4.0, the most recent major version, added three analysis tools at once.

impact_map draws the ripple effect of a single article as a graph. Ask for “cases citing Article 103 of the Civil Act” and it does a reverse traversal across Supreme Court rulings, Constitutional Court decisions, legal interpretations, administrative appeals, and local ordinances, then a forward traversal over the other statutes that article cites — and auto-generates mermaid graph code that renders directly in claude.ai.

time_travel diffs a statute across two points in time. Give it “Personal Information Protection Act 2020-01-01 vs 2025-11-01” and it pulls the text in force at each date, classifies additions (+), deletions (-), and changes (△) per article, and shows the before/after text plus character-count deltas. It is especially useful for frequently amended laws.

action_plan turns a citizen’s plain-language situation into a five-step guide. Type “I can’t get my rental deposit back” and it unfolds into STEP 1 situation diagnosis (auto-identifying the Housing Lease Protection Act) → STEP 2 rights and remedies (case law) → STEP 3 filing bodies and deadlines → STEP 4 required documents and forms → STEP 5 pitfalls and cautions, including a pointer to the Korea Legal Aid Corporation.

Operational details

The README’s release notes are refreshingly candid about the pains of running a remote server. In v3.3.0, the remote server hosted on fly.dev was periodically OOM-killed and restarted, invalidating session IDs — fixed by switching to MCP’s official stateless pattern, which builds a fresh Server + Transport per request and releases it immediately on completion. API keys are isolated per request with AsyncLocalStorage to prevent race conditions.

The v3.5.5 hotfix is just as telling. The government Open API began classifying Node.js’s default User-Agent (undici/...) as a bot and rejecting it, killing the tools across all cloud hosting. The error message — “please register the correct server IP address” — made it look like IP-whitelist blocking, but the real cause was UA inspection. A one-line patch injecting a normal browser UA into the default headers restored everything. Annex and form parsing is handled by kordoc, an engine by the same author that auto-converts HWPX, HWP, PDF, XLSX, and DOCX into Markdown.

Installation paths are plentiful: a one-line Claude Code plugin install, a claude.ai custom connector (https://korean-law-mcp.fly.dev/mcp?oc=...), a desktop-app config file, an npm global install, and direct CLI use — five in all. It is MIT-licensed and has over 1,700 stars on GitHub.

Insights

korean-law-mcp is interesting not because it is a legal search tool but because it is a record of empirically probing the right abstraction level for MCP tool design. v2’s “one API equals one tool” was intuitive but wasted the scarce resource of LLM context; v3 folded 89 tools into 17 with a domain enum and dispatch table while holding feature coverage at 100%. That runs exactly opposite to the REST API habit of slicing endpoints ever finer — and shows that when the consumer is an LLM rather than a human, the cost function of abstraction changes.

The second lesson is that anti-hallucination is not a single feature but a system-wide signaling problem. verify_citations exists to catch fake statutes, yet the tool itself produced false negatives, and more fundamentally other tools were causing hallucinations by not clearly signaling failure to the LLM. The [NOT_FOUND] machine markers, the bulk fix of isError flags, the removal of silent-drop patterns in the chain tools — all of it serves one goal: making a tool say “I don’t know” when it doesn’t. Anyone building an MCP server for a domain where accuracy is everything will hit this.

Third, the project shows the value of repackaging public data to be LLM-friendly. The legislation API was already free and open, but it only becomes usable from a one-line natural-language question once domain knowledge is layered on: automatic abbreviation resolution, article-number conversion between human and API forms, unified search across 17 ruling domains. The gap between a government agency opening an API and that API actually getting used was closed here, in open source, by a single civil servant. The same pattern — public API plus MCP wrapper plus domain knowledge — transfers cleanly to other public-data areas like tax, patents, and statistics.

References

Project

Model Context Protocol

Legal data sources

Background concepts

© 2024 - 2026 ICE-ICE-BEAR-BLOG
Built with Hugo
Theme Stack designed by Jimmy