Korean University Disclosure Indicators (대학알리미)

academyinfo-mcp

English | 한국어

academyinfo-mcp is an independent, read-only MCP server for factual queries and side-by-side comparison of Korean university disclosure indicators. It uses a bundled local snapshot: no API key and no runtime network connection are required.

This project is not affiliated with, endorsed by, approved by, sponsored by, or maintained by the Ministry of Education, KCUE, KEDI, data.go.kr, academyinfo.go.kr, or any university.

Quickstart

academyinfo-mcp runs with no install, no API key, and no login. It reads a bundled snapshot of Korean university disclosure indicators (대학알리미, KOGL Type 1) and answers factual queries and side-by-side comparisons from your AI assistant.

Claude Desktop

Add to claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/), then restart Claude Desktop:

{  "mcpServers": {    "academyinfo": {      "command": "npx",      "args": ["-y", "academyinfo-mcp"]    }  }}

Cursor

Add the same command/args shape to ~/.cursor/mcp.json (or a project .cursor/mcp.json), then reload the MCP server list.

Any MCP stdio client

npx -y academyinfo-mcp

The server speaks MCP over stdio (JSON-RPC on stdout, diagnostics on stderr).

Hosted endpoint (nothing installed)

For a client that cannot run a local process. A client that registers connectors by a URL field (for example claude.ai custom connectors) takes just the address:

https://academyinfo-mcp-433006350023.asia-northeast3.run.app/mcp

A client that takes a pasted mcpServers object takes this entry:

{  "mcpServers": {    "academyinfo": {      "description": "대학알리미 공시데이터 — 488개 대학·캠퍼스 17개 지표 비교",      "url": "https://academyinfo-mcp-433006350023.asia-northeast3.run.app/mcp"    }  }}

No credentials, so no headers. Same bundled snapshot and same eight tools as the local form.

This is a personal deployment offered as-is: it can go away without notice, and it is not covered by any availability guarantee (see Support). For work you need to reproduce, prefer the local form and pin a version — [email protected].

More on the pasted-object format, including the transport field that does not belong in it, is in examples/.

Try it

Ask your assistant something like:

"전남대와 부산대의 취업률과 경쟁률을 비교해줘."

The server never guesses an ambiguous name. "전남대학교" matches two campuses, so the first explore_universities call returns status: "ambiguous" with the candidates instead of a table:

전남대학교 → 2 candidates    전남대학교 / 본교       (국립, 광주)    전남대학교 / 제2캠퍼스  (국립, 전남)

Pick a campus and ask again (assistants usually do this follow-up on their own):

"전남대학교 본교와 부산대학교로 비교해줘."

The second call returns status: "ok" with each institution's values and the source, year, and unit attached — for example (2025 bundled snapshot):

This two-step flow is the designed behavior, not a failure: ambiguous names return candidate campuses instead of guessing, and comparisons never rank, score, or pick a winner — the numbers are presented as-is with provenance so you decide.

Evidence-scoped status

Implemented in this checkout:

• a file-first, offline, read-only stdio server backed by the bundled 15118998 derivative database;
• exactly eight registered MCP tools (the seven legacy tools plus explore_universities);
• a schema-validated, KOGL-attributed data-only indicator catalog at data/seed/indicators.json;
• Node engine >=22 <23 and the closed direct production dependency set @modelcontextprotocol/[email protected], [email protected], [email protected], and [email protected];
• ambiguity handling that returns candidates rather than guessing, and factual comparisons without scores, ranks, winners, or recommendations;
• a checkout-only stateless Streamable HTTP entry point (dist/src/http.js, POST-only, excluded from the npm package) alongside the packaged stdio entry point, with readOnlyHint/openWorldHint annotations on all eight tools.

Published: [email protected] is live on the public npm registry (latest), published by hand from an isolated terminal ceremony (docs/manual-publish-runbook.md). The initial 0.1.0 release was smoke-verified with an anonymous npx -y [email protected] install that resolves the exact registry tarball and lists all eight tools; later releases (0.1.1, 0.1.2, 0.2.0, 0.3.0, 0.3.1, 0.4.0) follow the same runbook.

Not yet performed (intentional, proportionate for a first solo release):

• the formal no-compile npx proof across all three official Node 22 lanes (macOS/arm64, Windows/x64, Ubuntu glibc/x64) — only a single-machine smoke test was run. Separately, a one-off manual run (2026-07-24, outside CI) on Ubuntu 24.04 glibc/x64 with Node 22 installed [email protected] anonymously, resolved better-sqlite3 from prebuilt binaries without compiling, and verified initialize, all eight tools, and an explore_universities call with JSON-RPC-only stdout; this is recorded observation, not the formal lane proof, and macOS/arm64 and Windows/x64 remain unverified;
• the receipt-bound candidate→client→promotion evidence chain and an actual-client (Claude Desktop) receipt;
• an approved unattended official download link or a completed annual refresh.

The package ships [email protected] as its sole backend. It installs from prebuilt binaries on the maintainer's Node 22 macOS/arm64 lane and, in the one-off manual run above, on Ubuntu 24.04 glibc/x64; the full three-lane prebuilt-only matrix has not been independently proven. Exactly one backend ships.

Requirements

Use Node >=22 <23. Node 24 and later are not supported or claimed. The public support matrix is limited to Node 22 on:

• macOS/arm64;
• Windows/x64;
• Ubuntu glibc/x64.

Those lanes are targets until public-install evidence is collected; local success is not public support evidence.

Local checkout use

The implemented local behavior can be exercised from a checkout:

npm installnpm run buildnode dist/src/index.js

The last command starts an MCP stdio server. It does not provide an interactive shell. MCP protocol output is written to stdout; diagnostics must remain on stderr.

No API-key environment variable is required. ACADEMYINFO_DB_PATH may select another compatible local database; otherwise the bundled seed is used. Runtime tools do not write the database or contact an external service. This development checkout defines npm run doctor, npm run refresh:acquire-validate, and npm run refresh:verify-artifact. Only doctor has its compiled program included in the packed npm artifact, where it is a local package/data/runtime diagnostic; it does not prove public installation, backend selection, refresh acceptance, or release approval. The compiled refresh programs, their TypeScript build sources, and required development dependencies are excluded from the package, so both refresh commands are checkout-only protected-workflow internals, not supported installed-package commands. Direct local invocation is never an authoritative acquisition, writer, or release receipt.

Remote endpoint (Streamable HTTP, checkout-only)

The npm package remains stdio-only. A checkout (or a container image built from it) also provides a stateless Streamable HTTP entry point for remote MCP clients such as claude.ai custom connectors:

npm run buildPORT=8080 ALLOWED_HOSTS=<public-host> npm run start:http

• POST /mcp accepts one JSON-RPC MCP request per call and answers a JSON response; GET (SSE streams) and DELETE (sessions) are not served because every request runs on a fresh server instance.
• GET /healthz and GET /health answer 200 ok for deployment health checks (on Cloud Run run.app domains the Google frontend consumes the literal /healthz path, so external probes must use /health).
• ALLOWED_HOSTS (comma-separated) enables DNS rebinding protection; leave it unset only behind a trusted proxy.
• dist/src/http.js is excluded from the npm package, so installed npx academyinfo-mcp behavior is unchanged.

A remote deployment serves the same bundled point-in-time snapshot under the same attribution and disclaimer boundaries: it is not a live feed and does not guarantee the latest data, and the operator remains unaffiliated with the Ministry of Education, KCUE, KEDI, data.go.kr, academyinfo.go.kr, or any university.

Version pinning

The Quickstart uses the unversioned academyinfo-mcp, which resolves to the current latest (now 0.4.0). To pin a specific version instead, use [email protected] in the args. Cursor and Codex use the same command/args shape as the Claude Desktop example; they are documented configurations and behave identically over MCP stdio.

Exact eight-tool scope

1. list_sources
2. list_indicators
3. search_university
4. get_university_metrics
5. compare_universities
6. explain_indicator
7. validate_source_coverage
8. explore_universities

The first seven contracts remain backward compatible. For client compatibility, tools/list registers explore_universities with this exact permissive Draft-07 outer input schema:

{  "$schema": "http://json-schema.org/draft-07/schema#",  "type": "object",  "properties": {    "university_queries": {},    "indicators": {}  },  "additionalProperties": {}}

The registered schema deliberately has no required keyword, and both property schemas plus additionalProperties are empty schemas. It is only the discovery boundary; the handler separately applies strict internal validation. The handler requires university_queries, rejects unknown top-level fields, accepts 1–10 unique trimmed query strings of 1–120 Unicode code points, and accepts an optional indicators array containing at most five unique supported nonempty indicator IDs. Invalid input fails before any journey is evaluated.

A valid explore_universities request is resolved against one local read-only snapshot. It is all-or-nothing. A missing or ambiguous university returns no partial metrics, comparisons, or explanations. Ambiguous matches return up to 20 repository-ordered candidates and never select one. The tool does not rank by metric values or substitute for a user's choice.

Every successful value or explanation retains source and license information, year or base year, unit, source column, derived/bundled status, and warnings. Missing source values remain explicit rather than being inferred.

Indicators and data boundary

The catalog contains exactly seventeen indicators from bundled dataset 15118998:

These dates describe the currently bundled point-in-time snapshot, not a live feed or latest-data guarantee. The catalog is JSON data, validated by a closed static schema and packaged under the data license; it is not generated executable source. The database, manifest, and catalog are independently cross-checked.

An exact 0 in a rate, ratio, or per-student indicator is not served as a value. It arrives in missing_metrics as zero_not_aggregatable with the source text preserved, because such a cell cannot be distinguished from a placeholder and would otherwise enter an average unnoticed. Headcount zeros are served with a note, since an institution can genuinely have none. The evidence and the open question are in docs/zero-values.md.

Dataset 15139279 is not bundled, enabled, sampled, normalized, or used for default employment results. Live OpenAPI access, scraping, and granular employment behavior are outside this release.

Refresh safety summary

Refresh approval is semantic, not based on a frozen checksum or fixed column count. Acquisition is read-only and separate from the fixed-path repository writer. A refresh must preserve indexed raw cells and prove:

• official source, workbook, and license identity;
• one unique identity mapping and one unique mapping for each of the seventeen logical indicators;
• fixed verified units and nondecreasing integer years;
• exact row coverage: every source row maps once, with seventeen numeric-or-missing classifications;
• missing values only for trimmed empty text or ASCII -;
• exact nonnegative decimal parsing and round-trip safety, with no rounding.

A post-download SHA-256 is integrity, change, and audit evidence only. A changed checksum, institution set, values, allowed missingness, unrelated columns, or a valid 24/26-column workbook can pass when all semantic invariants pass. A prior-checksum match does not authenticate a source. See docs/refresh-release-runbook.md.

Automated acquisition currently stops at DOWNLOAD_LINK_MISSING: the reviewed canonical page is reachable and its license verifies, but it exposes no anonymously fetchable download link, and policy forbids guessing an endpoint or scraping a replacement. Refresh is therefore performed manually — see docs/manual-refresh.md.

Freshness and release behavior

A refresh incident has an immutable first-seen time and a seven-day (604800000 ms) deadline; metadata, ETag, or repeated-failure drift does not reset it. Invalid official timestamps do not become trusted times.

• Equal reacquired bytes may close an incident only after origin/license/workbook validation and an administrator-attested verified-no-change receipt.
• Differing bytes enter the changed lifecycle and close only when the matching release-data digest is promoted to latest.
• Failed validation or release leaves the last-known-good data and public version in service.
• Rolling back reopens the original changed-event clock; a corrected release uses a new SemVer.

Candidate publication is not completion. Public matrix evidence and actual Claude Desktop/macOS evidence must be joined into protected receipts before administrator-approved promotion. Rollback restores the prior latest, deprecates the bad version, preserves evidence, and reopens the incident when applicable. Candidate publication, promotion, and rollback all serialize npm registry mutation through the academyinfo-mcp-registry-mutation concurrency group with cancel-in-progress: false. After promotion, the uploaded post-mutation promotion.v1.json must be retained and persisted byte-identically in a protected immutable default-branch evidence commit at evidence/releases/<version>/promotion.v1.json before rollback is available. After rollback, its uploaded post-mutation rollback.v1.json must likewise be retained and persisted byte-identically at evidence/releases/<bad_version>/rollback.v1.json.

Public-install proof boundary

Local installs and local tarballs do not satisfy public acceptance. Each official lane must use a fresh home, cache, working directory, and configuration outside the checkout; an explicit public registry; no reachable local artifact; and exact:

npx -y academyinfo-mcp@<version>

The proof must record the installed application and the complete direct production dependency set—@modelcontextprotocol/[email protected], [email protected], [email protected], and [email protected] in the current single-backend package—plus each registry integrity; candidate tag and provenance/signature evidence; Node/OS/architecture (and glibc on Ubuntu); active Python/node-gyp/compiler traps with a demonstrated canary; verbose sanitized install logs; initialization; the exact eight-tool list and explore_universities schema; a bundled-data query; no-key behavior; and JSON-RPC-only stdout. A separately approved backend replacement requires the verifier and this closed dependency set to change together. Promotion is prohibited if any lane compiles native code, resolves another direct dependency version, reaches a local package, or lacks the required evidence.

Administrator prerequisites

Before candidate publication, an administrator must establish npm package identity/history and ownership, select an unused SemVer, confirm 2FA/trusted-publishing and provenance readiness, configure exactly the protected environments refresh-pr-writer, npm-candidate, public-candidate-proof, claude-desktop-client-proof, npm-promotion, and npm-rollback with required approvals and artifact retention, define protected ACADEMYINFO_RELEASE_ADMINISTRATOR as the exact receipt-authority identity, and define ACADEMYINFO_PUBLIC_INSTALL_VERIFIER_SHA256 plus ACADEMYINFO_RELEASE_RECEIPT_VERIFIER_SHA256 from the reviewed current verifier policy. The administrator must resolve the single-backend gate and approve the immutable candidate, client-proof, promotion, and any rollback receipts at their separate privilege boundaries. Every protected verifier call binds the configured authority and reviewed verifier bytes; neither receipt-provided identity nor caller-selected historical code is authority. No repository document selects a version or grants those approvals.

Support

This project is published as-is, without support. Issues, questions, and feature requests are not monitored and will generally not receive a reply, and there is no service level for availability, data freshness, or response time. It is a personal open-source project maintained on a best-effort basis by one person and is not operated by any institution.

You are welcome to use it under the licenses below, and to fork it if you need changes. Do not depend on it for anything where an unanswered question or an unpatched defect would matter.

License and privacy boundaries

• Code is MIT licensed (LICENSE).
• Bundled 15118998 data is a normalized derivative under KOGL Type 1 attribution requirements (DATA_LICENSE.md, NOTICE.md, and data/seed/LICENSE.15118998.md). Code and data licenses remain separate.
• Release artifacts, receipts, logs, errors, examples, and MCP responses must not contain credentials, service keys, signed query strings, local user names, machine identifiers, or private filesystem paths.
• Raw workbooks and signed download URLs are not package contents.

The normative scope and gates are maintained in the repository at PROJECT_CHARTER.md, NON_GOALS.md, and RELEASE_CHECKLIST.md.