
medxpert-ortho-mcp
io.github.zhaoxinghua09-cell/medxpert-ortho-mcp
Orthopedic device testing MCP: lab match, quotes, UDI, standards, timelines, gap reports. Offline.
Documentation
medxpert-ortho-mcp · Orthopedic device testing connector
mcp-name: io.github.zhaoxinghua09-cell/medxpert-ortho-mcp
English first, Chinese below.
An MCP server with 8 tools for orthopedic device testing work: lab matching, quotation estimation, UDI checks, dual-temporal standards lookup, class/path judgment trees, region routes, registration timelines and gap reports. Offline, standard library only, zero dependencies.
It exists because four MedXpert expert packs (medxpert-standards, medical-device-reg-translator,
medical-device-testing-link, med-device-compliance-toolkit) had been referencing a connector called
medxpert-ortho-mcp that was never built. This is that connector, built.
What is real and what is sample
| Aspect | Status |
|---|---|
| The server, the 8 tools, the protocol layer | Real. Implemented, unit-tested, published to PyPI and to the official MCP Registry |
The datasets shipped in datasets/ | Sample. Invented labs, invented prices, invented UDI-DIs. Illustrative only |
| Every tool result | Carries sample_data: true until you set ORTHO_DATA_DIR |
The bundled data is not an accreditation directory, not a price commitment, and not
a regulatory conclusion. Point ORTHO_DATA_DIR at your own JSON files to use real evidence.
Data that does not fit fails loudly
A wrong data path used to be the most dangerous state this server could be in: the tools would keep
answering, from sample data, and sample_data: true was the only clue. That is a silent failure with
a regulatory conclusion on the far end of it, so it is now an error:
| Situation | Behaviour |
|---|---|
ORTHO_DATA_DIR set, directory missing | The call fails with the path in the message. It never falls back to samples |
| A required dataset file missing | Fails, naming the file and the directory in use |
| Malformed JSON | Fails, naming file, line and column |
| JSON valid but top-level keys wrong | Fails, listing the keys that are required |
| Structurally bad content (orphan lab reference, 13-digit UDI, unknown status, empty phase list) | Reported by the check below, per file and per cross-check |
Check your own data before relying on it:
ORTHO_DATA_DIR=/path/to/your/datasets python -m medxpert_ortho.server --validate-data
It prints a structured report, DATA CHECK: PASS or FAIL, and exits 1 on failure — so it can be
a CI step rather than a ritual.
Run
uvx medxpert-ortho-mcp # stdio MCP server
python -m medxpert_ortho.server --selftest # offline self-test, no host needed
python -m medxpert_ortho.server --validate-data # check ORTHO_DATA_DIR (or the bundled samples)
MCP host configuration:
{"mcpServers": {"medxpert-ortho-mcp": {"command": "uvx", "args": ["medxpert-ortho-mcp"]}}}
Custom data directory:
{"mcpServers": {"medxpert-ortho-mcp": {
"command": "uvx", "args": ["medxpert-ortho-mcp"],
"env": {"ORTHO_DATA_DIR": "/path/to/your/datasets"}}}}
Tools
| Tool | What it does | Honest limit |
|---|---|---|
match_labs | Ranks labs by capability keyword, region and device description | Sample data has 3 labs and a few capability items |
calc_quote | Price range by lab and scope, total estimated by quantity (>included samples +12%) | Prices are illustrative; not a commercial offer |
check_udi | 14-digit UDI-DI lookup, returns class / sterile / single-use | Misses are explained as sample-data scope, not proof of absence |
query_standard | Keyword search with as_of dual-temporal judgement (in force? superseded?) | Status fields are demo logic; verify against the official current version |
find_judgment_tree | Step-by-step classification / route questions | A demo tree, not a review decision |
get_region_path | Authority, route note and required documents per region | Sample routes only for CN / CN-JS / CN-SH |
check_timeline | Expands relative days into calendar dates, marks completed / in_progress / not_started, flags overdue | Durations are illustrative; official clocks start at acceptance |
report_gap | Compares present documents against the required checklist | Checklist is sample data |
Design notes
- Zero dependencies. JSON-RPC 2.0 over stdin/stdout, implemented with the standard library only. No network calls at runtime; nothing leaves the machine.
- Provenance in every answer. A tool that cannot say whether it answered from sample data or from your own data is a tool you cannot audit, so every result says which.
- Mismatch is reported, not smoothed over. A UDI miss says "sample scope", a tree miss lists what trees exist, an empty lab match returns a hint rather than an empty success.
- Bad input is an error, not a quiet fallback. See above: the server will not answer from samples
while your data path is broken. The check is also runnable as
--validate-datawith a non-zero exit.
Gate
.github/workflows/ci.yml runs the unit tests and tools/version_gate.py, which checks
version consistency across pyproject.toml, __init__.py, server.json and the declared tool
count, and reports how many uses: references are pinned to full SHAs.
medxpert-ortho-mcp · 骨科器械检测对接连接器
先英后中。
一个 8 工具的 MCP server,用于骨科器械检测对接:机构匹配、报价估算、UDI 校验、双时态标准检索、 分类/路径判定树、地区路径、注册时间轴、缺口报告。离线、仅标准库、零依赖。
它存在的理由:四个 MedXpert 专家包(medxpert-standards、medical-device-reg-translator、
medical-device-testing-link、med-device-compliance-toolkit)从 2026 年起一直引用一个叫
medxpert-ortho-mcp 的连接器,而那个连接器从未被造出来。这就是它,现在有了。
哪些是真的,哪些是示例
| 方面 | 状态 |
|---|---|
| 服务本身、8 个工具、协议层 | 真的。 已实现、有单元测试、已发布到 PyPI 与官方 MCP Registry |
datasets/ 里的数据 | 示例。 虚构机构、虚构价格、虚构 UDI-DI,仅供演示 |
| 每个工具返回值 | 只要没设 ORTHO_DATA_DIR,一律带 sample_data: true |
包内数据不是资质目录、不是报价承诺、不是法规结论。要用于真实工作,把
ORTHO_DATA_DIR 指向你自己的 JSON 数据目录。
数据不合规就报错,不静默降级
数据路径写错,曾是这台服务最危险的状态:工具照常作答,用示例数据答,唯一线索只有一个
sample_data: true。这是会一路传导到法规结论的静默失败,所以现在改成报错:
| 情形 | 行为 |
|---|---|
ORTHO_DATA_DIR 设了但目录不存在 | 调用直接失败,错误信息里带路径;绝不退回示例数据 |
| 必需的数据集文件缺失 | 失败,并指出文件名与当前数据目录 |
| JSON 语法错 | 失败,并指出文件、行、列 |
| JSON 合法但顶层键不对 | 失败,并列出必需的顶层键 |
| 结构性问题(报价指向不存在的机构、13 位 UDI、非法 status、阶段为空) | 由下面的体检逐文件、逐交叉项报出 |
用自己的数据之前先体检:
ORTHO_DATA_DIR=/path/to/your/datasets python -m medxpert_ortho.server --validate-data
输出结构化报告,末行 DATA CHECK: PASS 或 FAIL,失败时退出码 1——因此它可以进 CI,
而不只是一次仪式。
运行
uvx medxpert-ortho-mcp # 起 stdio MCP 服务
python -m medxpert_ortho.server --selftest # 离线自测,不需要宿主
python -m medxpert_ortho.server --validate-data # 体检 ORTHO_DATA_DIR(或包内示例)
工具与诚实的边界
| 工具 | 做什么 | 边界 |
|---|---|---|
match_labs | 按能力关键词、地区、器械描述给机构打分 | 示例数据仅 3 家机构、少数能力项 |
calc_quote | 按机构与能力项给价格区间,超出含样量 +12% | 价格为演示值,不构成商务报价 |
check_udi | 14 位 UDI-DI 查表,回出分类/无菌/一次性 | 未命中会说明是示例范围,不是"不存在"的证据 |
query_standard | 关键词检索 + as_of 双时态判定(是否生效、是否被替代) | 状态字段为演示逻辑,以官方现行版为准 |
find_judgment_tree | 逐步问答式分类/路径判定 | 演示结构,不是审评结论 |
get_region_path | 受理机关、路径说明、必备材料 | 仅 CN / CN-JS / CN-SH 三条示例路径 |
check_timeline | 相对天数展开成日历,标 completed / in_progress / not_started 与逾期 | 天数为示例;官方时限自受理起算 |
report_gap | 已具备材料对照必备清单,报缺口与完整度 | 清单为示例数据 |
门禁
ci.yml 跑单元测试 + tools/version_gate.py:核对 pyproject.toml、__init__.py、server.json
与声明的工具数是否一致,并自报 uses: 有多少处钉了完整 SHA。
medxpert-ortho-mcppip install medxpert-ortho-mcpRelated Servers
io.github.zhaoxinghua09-cell/medxpert-library-connector
Local RAG MCP server for your own .md knowledge base: BM25 + local Ollama, offline, zero egress.
io.github.zhaoxinghua09-cell/medxpert-reg-connector
Read-only MCP for medical-device regulation (NMPA/FDA/MDR/PMDA): 27 hubs, 7 tools, freshness guard.
io.github.zhaoxinghua09-cell/medxpert-standards-connector
Read-only local MCP index of medical-device standards (ISO/GB/YY/ASTM/IEC) with official links.