Managing Your Warehouse with AI Agents: A Practical Guide to MCP in Flash Warehouse WMS

When AI Agents Meet Warehouse Management

The traditional workflow for warehouse management is rigid: open a browser, log into the backend, click through menus, fill out forms, submit. For high-frequency daily operations -- checking stock, creating purchase orders, approving receipts -- this process is repetitive and time-consuming.

What if a warehouse manager could simply tell an AI "check the stock for Coca-Cola" and get precise inventory data back in seconds? Flash Warehouse WMS makes this possible through its implementation of the Model Context Protocol (MCP). To our knowledge, Flash Warehouse is the first WMS in China to support MCP.

What Is MCP

MCP (Model Context Protocol) is an open protocol published by Anthropic in 2024 that defines a standardized communication layer between AI models and external tools. In practical terms, MCP lets an AI Agent call functions on external systems -- reading data, executing writes, and triggering workflows -- through a structured JSON-RPC interface over stdio.

The protocol works through three key mechanisms:

• Tool registration: The server declares its capabilities (tools) to the AI, each with a name, parameter schema, and description
• Standardized invocation: The AI calls these tools via JSON-RPC messages and receives structured return values
• stdio transport: Communication happens over standard input/output, making it a natural fit for CLI tool integration

AI clients with MCP support include Claude Code (Anthropic's official CLI), Cursor, and Windsurf.

fwh: Flash Warehouse's CLI and MCP Server

Flash Warehouse's MCP capability is provided by fwh (Flash Warehouse CLI), a static binary written in Go with zero runtime dependencies. It runs on macOS, Linux, and Windows.

fwh serves two roles simultaneously:

1. Traditional CLI tool: 23 top-level commands for direct terminal-based warehouse operations
2. MCP server: Started via fwh mcp, exposing 44 read-only tools and 106 total tools (with writes enabled)

Configuring Claude Code

Add the following to ~/.claude/mcp.json:

{
  "mcpServers": {
    "flash-warehouse": {
      "command": "/absolute/path/to/fwh",
      "args": ["mcp"]
    }
  }
}

To enable write operations (creating bills, approvals, etc.), change args to:

"args": ["mcp", "--enable-writes"]

The configuration format is identical for Cursor and Windsurf.

Real Usage Scenarios

Once configured, you can operate your warehouse using natural language in an AI conversation:

• "Check stock for Coca-Cola" -- The AI calls goods_list with a fuzzy name search and returns matching products with quantities and warehouse locations
• "What are today's sales?" -- The AI calls bi_today_sales_revenue and returns revenue and order count
• "Create a purchase receipt for 500 units of item 6954767410388" -- The AI calls purchase_input_create with the specified goods data
• "Approve bill INQ-20260411-001" -- The AI calls bill_audit to complete the approval workflow
• "Which products are below their alert thresholds?" -- The AI calls alert_overview and returns an inventory alert summary

Every tool returns structured JSON data that the AI parses and presents in natural language.

The Three-Layer Safety Model

Letting an AI Agent operate on production data demands rigorous safety guarantees. fwh is built around three hard rules:

Layer 1: Login-First

The MCP server refuses to start without an active login session. You must run fwh login in your terminal first. This means an AI Agent can never log in on its own -- login credentials never appear in conversation history.

fwh login --user 13800138000
# Password: ********
# logged in as 13800138000 (user_id=1)

Layer 2: Tenant-Locked

binding_user_id is the core field that isolates multi-tenant data in Flash Warehouse WMS. In fwh, this value is captured once at login time from the server response and stored in an unexported Go struct field with a getter and no setter. No CLI flag, environment variable, or MCP tool argument can override it.

A source-level audit test (TestTenantLock_UserIDChokepoints) verifies on every build that session.UserID() is read in exactly 3 fixed locations within api/client.go. Any new read site outside these chokepoints fails the test.

Layer 3: Write Opt-In

fwh mcp registers only read-only tools (44) by default. Write tools (create, modify, delete, approve) are registered only when --enable-writes is explicitly passed. An unregistered tool does not exist at the MCP protocol level -- the AI cannot call a tool that is not in the server's registry.

Additionally, every write tool's description begins with WARNING: to ensure the AI's reasoning surfaces the destructive nature of the operation before execution. For example:

WARNING: APPROVE a bill. Commits inventory changes, cannot be undone.

Tool Coverage

The MCP tools in fwh cover the full functionality of Flash Warehouse's PC management frontend:

Comparison with Traditional Workflows

Consider a common daily task: "check alerts and restock."

Traditional approach: Open browser -> Log in -> Navigate to inventory alerts -> Identify low-stock items -> Switch to purchase module -> Fill out purchase form -> Submit. Approximately 5-10 minutes.

MCP approach: In Claude Code, say "Check inventory alerts and create a purchase receipt for items below threshold." The AI sequentially calls alert_overview -> alert_list -> purchase_input_create. Approximately 30 seconds.

Getting Started

# 1. Build fwh (requires Go 1.22+)
cd flash_warehouse_cli
go build -o ./bin/fwh ./cmd/fwh

# 2. Log in
./bin/fwh login --user your-phone-number

# 3. Verify identity
./bin/fwh whoami

# 4. Configure Claude Code (edit ~/.claude/mcp.json)
# 5. Restart Claude Code and start managing your warehouse with natural language

Conclusion

The MCP protocol introduces a fundamentally new interaction paradigm for warehouse management. Flash Warehouse WMS, through fwh's 44 read-only and 106 total MCP tools, achieves complete feature parity with the PC frontend while maintaining production safety through its login-first, tenant-locked, write-opt-in security model.

Flash Warehouse WMS is completely free. Visit flashwarehouse.cn to learn more.