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.

Related MCP server: fix-mcp

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.Pool for 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

1. Create a Postgres database and set DATABASE_URL in .env (copy .env.example).

2. Load the schema and synthetic data:

   npm run db:setup

3. Install dependencies and start the server:

   npm install
   npm run dev

4. 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)

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.