inventory-ops-mcp
A remotely hosted MCP server that lets ops personnel investigate and resolve inventory oversells without engineer involvement.
README
inventory-ops-mcp
A remotely hosted MCP server that lets an ops person investigate and resolve inventory oversells without engineer involvement.
Workflow this solves
Oversell: a SKU shows more stock reserved across open orders than it
actually has (available = total_stock - active_reservations goes
negative). Root cause is usually one of two things:
- A stale reservation — an order was cancelled or expired, but its stock reservation was never released. Safe to auto-fix.
- A genuine conflict between paid/open orders. Not safe to auto-fix — gets escalated to a human with an audit trail instead.
Project structure
db/
schema.sql tables, enums, indexes
seed.sql synthetic data covering all 4 demo scenarios
src/
config.ts env config
db/
pool.ts single shared pg.Pool for the whole process
queries/ raw SQL, grouped by domain (inventory, audit)
tools/
inventory/ one file per tool + an index.ts that registers them
index.ts top-level registry - add new categories here later
server/
createMcpServer.ts builds one McpServer instance (per session)
httpServer.ts Streamable HTTP transport, session handling
index.ts entrypoint
Tools are grouped by domain category (tools/inventory/) rather than dumped
in one file, so a second category (e.g. tools/orders/) can be added later
without touching existing code — just create the folder, register its tools
in its own index.ts, and add one line to src/tools/index.ts.
Architecture notes
- One MCP server instance per client session (Streamable HTTP, session-id based) — this is the SDK's own recommended pattern, since protocol/session state can't be safely shared across clients.
- One shared
pg.Poolfor the entire process — imported by the query modules, not created per session. Sessions are isolated at the protocol layer; the database connection is not. - The resolution tool (
release_stale_reservation) enforces its safety boundary in code, not by trusting the model: it refuses anything that isn't a cancelled/expired order's reservation, regardless of what the AI requests.
Setup
- Create a Postgres database and set
DATABASE_URLin.env(copy.env.example). - Load the schema and synthetic data:
npm run db:setup - Install dependencies and start the server:
npm install npm run dev - Point an MCP client at
http://localhost:3000/mcp(or your deployed Render URL). For Claude Code:claude mcp add --transport http inventory-ops http://localhost:3000/mcp
Demo scenarios (seeded)
| SKU | Situation | Expected tool path |
|---|---|---|
| SKU-BULK-1 | No issue | get_stock_and_reservations shows no oversell |
| SKU-BULK-PAID-2 | Oversold by paid + open orders | must escalate — release refused |
| SKU-STALE-CANCEL-3 | Oversold by a stale reservation (cancelled order) | release_stale_reservation resolves it |
| SKU-STALE-EXPIRED-4 | Same, order expired instead of cancelled | release_stale_reservation resolves it |
Out of scope (explicit)
- Stale physical-count reconciliation (a related but separate workflow) — not built in this pass; escalation for that case is created but not auto-resolved.
- Authentication, a frontend, and CI/CD — excluded per the assignment brief.
- Multi-SKU orders (one line item per order, to keep the schema simple).
Verification
See AI-WORKLOG.md (to be added) for how each tool's behavior was verified
against the seeded scenarios above.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。