huiwen-mcp

Model Context Protocol (MCP) Server for the Huiwen Library Management System (Libsys / OPAC) — An AI-oriented read-only data gateway for libraries: Enables AI clients like Claude / Cherry Studio / DeepSeek to securely and audibly retrieve catalog records, item availability, circulation statistics, and union catalogs.

An official adaptation layer developed by a university library, adhering to the security baseline of read-only by default, least privilege, and full-chain auditing.

• Protocol: Model Context Protocol (Anthropic open standard, same technical approach as Yale Library's catalog integration)

• Runtime: Python ≥ 3.10 · FastMCP 3.x

• Data Sources: demo (zero-dependency demonstration) / opac (Huiwen OPAC public web protocol) / oracle (Huiwen Libsys database read-only direct connection)

• License: Apache-2.0 (recommended, see License & Compliance)

Table of Contents

1. Features

2. System Design Philosophy

3. Implementation Technical Approach

4. Quick Start

5. Configuration (Environment Variables / .env)

6. Tool List

7. Client Integration Examples

8. Use Cases

9. Security & Compliance

10. Testing

11. Project Structure

12. Roadmap

13. License & Compliance

14. Troubleshooting

Related MCP server: dms-mcp-server

Features

Design Trade-off: Write operations (renewal, reservation, interlibrary loan ordering) are intentionally not implemented — this project only provides "secure and auditable reading"; all write paths are delegated to the original business systems and manual processes.

System Design Philosophy

Positioning: Data Gateway / Skill Layer, Not a Database Proxy

AI clients (large language models) never directly connect to the Huiwen database. All queries are encapsulated through a controlled tool layer:

┌─────────────── AI 客户端（Claude / Cherry Studio / 自研 Agent / 本地 LLM） ───────────────┐
│                                      │                                                    │
│             stdio（子进程协议）        │        Streamable HTTP（服务化 / 网关 / SSO）        │
└──────────────────────────────────────┼────────────────────────────────────────────────────┘
                                       ▼
┌───────────────────────────────────────────────────────────────────────────────────────┐
│   huiwen-mcp（FastMCP 3.x）                                                            │
│   ┌─────────────── 安全链 _guard ───────────────┐                                        │
│   │ 认证(Auth) → 限流(TokenBucket) → 门控(PII/读者) │   ← 每个工具必经                     │
│   └──────────────────────────────────────────────┘                                        │
│   │ 工具层：search_books / get_book_detail / union_search / get_reader_* / … (12 个)     │
│   └──────────────────────────────────┬───────────────────────────────────────────────────┘
                                       ▼
┌───────────────────────────────────────────────────────────────────────────────────────┐
│   适配器（可插拔数据源，统一 CatalogBackend 接口）                                       │
│   ├─ OracleBackend：白名单参数化 SQL（db/queries.py 封闭集）   → 汇文 Libsys 只读账号    │
│   ├─ OpacBackend：白名单参数调汇文 OPAC 公开网页协议           → opac 站点                │
│   └─ DemoBackend：内置样例数据                                 → 离线演示/测试            │
└───────────────────────────────────────────────────────────────────────────────────────┘

• Single Responsibility per Layer: The adapter is only responsible for data retrieval; _guard is only responsible for security; auditing independently writes JSONL logs; the upper-layer AI only interacts with tool signatures, unaware of backend differences (three backends, same signatures).

• Secure by Default: data_source=demo runs with zero dependencies; opac/oracle require explicit configuration; patron-sensitive tools require an admin token; write operations are disabled by default; external union services are disabled by default.

Why MCP

• MCP is an open standard for connecting AI to "databases/business systems" (Anthropic released Nov 2024, ecosystem includes GitHub/cloud vendors/database vendors). Choosing an open standard over a private API ensures: client replaceability (Claude/Cherry Studio/DeepSeek/custom Agent), service reusability across multiple systems, and long-term freedom from vendor lock-in — this is the same path Yale Library took with MCP for catalog integration.

• FastMCP provides stdio / HTTP dual transport for server implementation, supporting both in-process and service-oriented deployment with a single codebase.

Transport Mode Selection: stdio vs HTTP

• stdio: Launched in-process with the client, zero operations overhead, lowest latency, suitable for personal/single-machine integration with AI desktop clients.

• HTTP (Streamable HTTP): Independent service, suitable for multi-user/centralized deployment; can be fronted by an OAuth2/JWT reverse proxy and campus unified identity for centralized auditing.

Implementation Technical Approach

Key Contracts (All Verified Against Live Systems)

• OPAC: Search results <ol id="search_book_list"> → <li class="book_list_info">, title/call number/holding copies/available copies/hit count; detail page item table; hot list.

• Union PROCAT: POST (GET→405); authentication via query parameter tk= (JWT issued by OPAC patron session getReaderJwt); items[].logic="1"(AND)/"2"(OR); field mapping any/title/author/subject/isbn/clcNumber/publisher/series. See docs/union-catalog-search.md for details.

⚠️ OPAC / Union are vendor-closed or third-party systems; contracts may change with deployment versions. All integration documentation is based on "live system verification" and validated using tests/test_*_live.py.

Quick Start

1) Installation

git clone <your-repo-url> && cd huiwen-mcp
# 方式 A：uv（推荐）
uv sync
# 方式 B：pip
python -m venv .venv
. .venv/bin/activate
pip install -e .

2) Run with Zero Configuration (demo data source, offline)

HUIWEN_DATA_SOURCE=demo uv run huiwen-mcp        # stdio 模式
HUIWEN_DATA_SOURCE=demo HUIWEN_TRANSPORT=http uv run huiwen-mcp   # HTTP 模式

demo includes sample bibliographic/patron data, suitable for smoke testing, development testing, and integration teaching.

2b) One-Click Docker Deployment

docker build -t huiwen-mcp:latest .
docker run --rm -it -e HUIWEN_DATA_SOURCE=demo huiwen-mcp:latest   # stdio，离线可跑

# 服务化（HTTP + 认证 + 审计）
docker run -d --name huiwen -p 8765:8765 \
  -e HUIWEN_TRANSPORT=http -e HUIWEN_DATA_SOURCE=opac \
  -e HUIWEN_OPAC_BASE_URL=https://opac.example.edu.cn \
  -e HUIWEN_AUTH_ENABLED=true -e HUIWEN_AUTH_BEARER_TOKEN=<强随机> \
  -v huiwen-audit:/var/log/huiwen huiwen-mcp:latest

More (Oracle 11g thick / compose / reverse proxy-level authentication with campus CAS) in docs/deployment-guide.md.

3) Connect to Real Data Sources (opac / oracle)

Copy .env.example to .env and fill in the values (.env is git-ignored):

cp .env.example .env
# 编辑 .env：设置 HUIWEN_DATA_SOURCE 与对应凭据
HUIWEN_DATA_SOURCE=opac
HUIWEN_OPAC_BASE_URL=https://opac.example.edu.cn      # 你们学校 OPAC 地址

Or use config.local.json (sensitive configuration auto-loaded, not committed to repository).

Configuration (Environment Variables / .env)

All configurations can be injected via environment variables (prefix HUIWEN_), and also support .env files (auto-loaded). Priority: Environment Variables > Explicit config.json / CONFIG_PATH > config.local.json Auto-Merge > Built-in Defaults.

General

OPAC

Oracle

Security

Tool List

For functional descriptions and integration assessment of Huiwen ACS/SIP2 interface services, see docs/huiwen-acs-sip2-interface-description-and-integration-assessment.md (authoritative field mapping, read-only subset candidates, explicitly prohibited items).

Patron tools are masked by default (include_pii=false does not return ID numbers/contact information; true requires admin).

Client Integration Examples

Claude Desktop / MCP-Supported Desktop Clients

{
  "mcpServers": {
    "huiwen": {
      "command": "/path/to/uv",
      "args": ["--directory", "/path/to/huiwen-mcp", "run", "huiwen-mcp"],
      "env": { "HUIWEN_DATA_SOURCE": "demo" }
    }
  }
}

Remote HTTP (Requires Authentication Gateway Setup)

HUIWEN_TRANSPORT=http HUIWEN_HOST=0.0.0.0 HUIWEN_PORT=8765 uv run huiwen-mcp

Clients use ${MCP_SERVER_URL} to connect to http://<host>:8765/mcp/ (Streamable HTTP). When HUIWEN_AUTH_ENABLED=true is enabled, the token is passed as a tool parameter token with each call; the HTTP Authorization header is not consumed by the server (see deployment guide §3.2).

Use Cases

For complete suggestions (including locally deployed LLM + RAG layered approach and domestic/international benchmarking), see docs/service-and-application-suggestions.md.

Security & Compliance

1. Read-Only by Default: All tools are read-only; write operations (renewal/reservation/interlibrary loan ordering) are intentionally not implemented.

2. Whitelist SQL: The Oracle backend only executes the parameterized SQL closed set within db/queries.py, no free SQL.

3. Full-Chain Gating: Authentication → Rate Limiting → Patron/PII Gating → Auditing (JSONL). Patron personal data requires an admin token and is masked by default.

4. Authentication Contract (Verified Against Live System): The token is passed via the tool parameter token (an optional parameter for each tool; _guard extracts it from the parameters and compares it with HUIWEN_AUTH_BEARER_TOKEN). The HTTP Authorization header passthrough is not implemented — transport layer TLS/unified identity is handled by the reverse proxy gateway; huiwen-mcp's own authentication is the second line of defense behind the gateway. Tokens are not written to audit logs (_guard pops them before recording).

5. Secrets Not Committed to Repository: DSN/password/JWT/site URLs are only passed via environment variables or config.local.json (git-ignored). The repository contains no real deployment data (see NOTICE).

6. External Services Handled with Caution: The Union PROCAT is a third-party multi-tenant system, disabled by default; confirm authorization with the union/service provider before enabling. OPAC is closed-source, has had public vulnerabilities historically; the adapter only uses whitelist parameters.

7. Vulnerability reporting and handling are in SECURITY.md.

Testing

Live database/live system tests are disabled by default (require local explicit setting of HUIWEN_LIVE_* to execute) to avoid touching any real systems. Docker images are not built/published by default (the release strategy is "publish only source code and documentation"): if an image is needed, please docker build locally (add --build-arg WITH_INSTANT_CLIENT=true for Oracle thick mode).

Project Structure

huiwen-mcp/
├── src/huiwen_mcp/
│   ├── server.py            # FastMCP 装配、stdio/http 启动、main()
│   ├── config.py            # 配置：env/.env/config.local.json 分层合并
│   ├── audit.py             # JSONL 审计
│   ├── adapters/
│   │   ├── base.py          # CatalogBackend 抽象
│   │   ├── demo.py          # 内置演示数据
│   │   ├── opac.py          # 汇文 OPAC 网页协议（含 union_search）
│   │   └── oracle.py        # Libsys 数据库只读（thin/thick）
│   ├── db/queries.py        # 白名单参数化 SQL（Oracle 后端唯一 SQL 来源）
│   ├── models/schemas.py    # pydantic 结果模型
│   └── tools/catalog.py     # 12 个 MCP 工具 + _guard 安全链
├── docs/                    # 表结构 / 联盟契约 / 服务与应用建议 / 部署指南 / SIP2 评估
├── tests/                   # demo/stdio/oracle-live/union-live
├── Dockerfile / compose.yaml / .dockerignore
├── .env.example / config.example.json / config.local.json(忽略)
├── LICENSE / NOTICE / SECURITY.md / CONTRIBUTING.md / CODE_OF_CONDUCT.md
└── pyproject.toml

Roadmap

• Phase 1: Read-only search MCP (demo + opac + oracle three backends)

• Phase 2: OPAC / Oracle live database integration testing, Union catalog integration testing (contract verified + token scheme)

• Phase 2 Remaining: Docker image (non-root, reproducible build) + deployment guide (including reverse proxy-level authentication template)

• Released: v1.0.0 tag + GitHub Release (source code and documentation; no CI/workflows, Docker images not automatically built)

• OAuth2/JWT gateway implementation for campus CAS / unified service portal (template ready, requires on-site configuration)

• Phase 2.5/3 Candidate: Huiwen ACS/SIP2 read-only subset (assessment in docs/huiwen-acs-sip2-interface-description-and-integration-assessment.md)

• Phase 3: RAG vector database + local LLM for intelligent book recommendations / reference consultation (see docs/service-and-application-suggestions.md)

• Phase 4: Huiwen new-generation platform OpenAPI integration

License & Compliance

Open Source License Version Suggestion

This project recommends using the Apache License 2.0 (the repository includes the complete LICENSE):

1. Permissive: Allows universities, vendors, and cloud platforms to freely use/modify/redistribute (including commercial use) as long as the copyright and license notice are retained—facilitating adoption by AI toolchains and third-party systems.

2. Patent grant: Apache-2.0 explicitly grants a patent license from contributors (Section 3), making it clearer and more "anti-litigation" when multiple institutions/parties (multiple universities, technology vendors) contribute jointly.

3. Contributor terms standardization: Implicitly grants a project license (Section 5, Contribution Grant), eliminating the need for each contributor to sign a separate CLA, consistent with common GitHub open-source project practices.

4. Differentiation: Compared to MIT, Apache-2.0 is more suitable for infrastructure-type projects formally released by institutions and potentially maintained by multiple parties over the long term.

If your library prefers a "minimalist style," you can revert to MIT at any time: simply replace the full LICENSE file, change license in pyproject.toml back to { text = "MIT" }, and update this section in the README.

Compliance Statement (Important)

• Contains no vendor/third-party source code: This project is an independent interoperability layer for the closed-source Huiwen/Libsys. It does not contain any proprietary code from Huiwen or the consortium. The OPAC/consortium contract is based solely on public web protocols and real-site response records. See NOTICE.

• No deployment-sensitive data is published with the repository: Actual DSNs, account passwords, OPAC login instances, consortium JWTs, reader PII, or vendor SECRET_KEY are not in the repository (SECURITY.md/CONTRIBUTING.md has set red lines strictly prohibiting any suspected sensitive data from being committed).

• Trademarks: Huiwen, Libsys, and OPAC are trademarks/product names of their respective owners (e.g., Jiangsu Huiwen Software). This repository uses them only for interoperability reference, not implying endorsement or affiliation.

• Before using this software, please confirm the authorization and usage boundaries with Huiwen Software, the consortium service provider, and your library's IT center.

Troubleshooting