# MCP use case: customer-service lookups > Staff ask Claude "where is order 1007?" or "do we have the Trailblazer tent?" and get the facts from the back office in one or two read-only tool calls. In the field test, Claude Opus 5.5 answered all 32 lookup runs correctly, with 2.0 tool calls per request on average; the tools spent 4.4 ms per request in total, the rest of the time was the model. **Cargo feature:** [`ai`](../build-features.md) · **Example:** [`examples/agents/mcp-shop`](https://github.com/pmsudhi/cameleer/tree/main/examples/agents/mcp-shop) · **Part of:** [Cameleer MCP](../mcp.md) ## What people ask, and what happens Requests typed into Claude with the shop's MCP server connected (sample date: Saturday 7 November 2026). The tool calls are the ones Claude made in the recorded run. | request | tool calls | what Claude answered | |---|---|---| | Where is order 1007? | `track_shipment` | with DPD, tracking DPD15501007, every scan, due tomorrow, Sunday 8 November | | where is ordr 1007 lol | `track_shipment` | the same answer: the typo and the tone do not matter | | Do we have the Trailblazer 2 tent in stock? | `search_products` | out of stock; 20 arriving on 14 November | | What's the email address on file for Priya Natarajan? | `search_customers` | priya.n@example.com, and that she is the only customer with that name | | Show me everything Marco Rossi has ordered this year. | `search_customers`, `list_orders`, `get_order` ×3 | three orders, £569 in total, in a table | | Which orders are still waiting to ship? | `list_orders` ×2 | five paid orders, oldest first, and one more waiting for payment | | Is order 1013 delivered yet? The customer says it never arrived. | `get_order`, `track_shipment`, `read_document` | delivered but left in the porch without a signature, and the shipping policy's next steps | | How much did we sell last week, and what was the top product? | `sales_report` | £1,479 from 4 orders, the top product, and which week it counted | | Where's order 9999? | `track_shipment` | it does not exist; asks for another way to find the customer | ## A recorded conversation **User:** Where is order 1007? Claude called `track_shipment` with `{"orderId": 1007}` and got the shipment and its events (2.9 ms in the server). It answered: > Order 1007 has shipped with DPD and is on its way. It's due to arrive tomorrow, Sunday > 8 November. Tracking number: DPD15501007. 4 Nov, 17:20: DPD collected it from our warehouse > in Swindon. [...] Two turns, one tool call. ## How the tool is built `track_shipment` is a route. `direct:load-order` (shared by several tools) reads the order into headers; the route then answers with the shipment and its tracking events, or with a readable error: ```yaml - route: id: track-shipment from: uri: ai-tool:track_shipment parameters: title: Track a shipment description: >- Where an order is: carrier, tracking number, delivery status, expected delivery date and the tracking events so far. Orders that have not shipped yet have no shipment. tags: acme parameter.orderId: integer parameter.orderId.description: The order number, such as 1007 parameter.orderId.required: true readOnlyHint: true idempotentHint: true steps: - to: direct:load-order - choice: when: - simple: "${header.order} == null" steps: - setHeader: name: CamelAiToolIsError constant: true - setBody: simple: "Order ${header.orderId} does not exist. Order numbers run from 1001; ask the customer to check the number." - simple: "${header.orderStatus} in 'pending,paid,cancelled'" steps: - setBody: simple: '{"orderId": ${header.orderId}, "orderStatus": "${header.orderStatus}", "shipment": null, "note": "Order ${header.orderId} has not shipped (status ${header.orderStatus})."}' otherwise: steps: - to: "sql:select carrier, tracking, status, eta from shipments where order_id = :#orderId?dataSource=jdbc:rustsql:mem:acme&outputType=SelectOne" - marshal: json: {} - setHeader: name: shipment simple: "${bodyAs(String)}" - to: "sql:select at, description from shipment_events where order_id = :#orderId order by at?dataSource=jdbc:rustsql:mem:acme" - marshal: json: {} - setBody: simple: '{"orderId": ${header.orderId}, "orderStatus": "${header.orderStatus}", "deliveredOn": "${header.orderDelivered}", "shipment": ${header.shipment}, "events": ${body}}' ``` What makes lookups work well with a model: - **`readOnlyHint` and `idempotentHint`** tell the client the tool changes nothing, so it can call it without asking the user. - **The description says what comes back and what to use next.** Claude picks `track_shipment` for "where is" and `get_order` for "what's in". - **Return the facts the next question needs.** `search_products` includes the incoming stock and its date, so "do we have it?" is one call, not two. - **Refuse in words.** For an unknown order the route sets `CamelAiToolIsError` and says what to do instead; the model passes that on rather than guessing. Exceptions are hidden from clients by default, so a readable refusal is the route's job. ## Try it ```bash cd examples/agents/mcp-shop cameleer run shop.camel.yaml --properties application.properties ``` Then connect Claude Desktop, the MCP Inspector or any MCP client to `http://127.0.0.1:18190/mcp`, or use stdio as the example's [README](https://github.com/pmsudhi/cameleer/tree/main/examples/agents/mcp-shop) shows. The test `shop_use_cases` in `crates/cameleer-main/tests/mcp_shop.rs` makes these calls over MCP and checks the answers. ## Related - [Questions that take several steps](multi-step-questions.md) - [Cameleer MCP](../mcp.md), [MCP server](../mcp-server.md), [ai-tool](../../components/ai-tool.md) ## Found a problem? If something on this page doesn't match what Cameleer does, please [open an issue](https://github.com/pmsudhi/cameleer/issues/new), or fork the repository, fix it, add a test, run `cargo test --workspace` and open a pull request.