Plug in a tool without rewriting chat

The point of the bridge is a boring change. You register a new MCP tool. You do not touch the SwiftUI chat view or the respond call.

let lookup = MCPTool(
    name: "lookup_order",
    description: "Return the local order status for an id the user already has.",
    inputSchema: [
        "type": "object",
        "properties": [
            "orderId": ["type": "string", "description": "Order id from the user"]
        ],
        "required": ["orderId"]
    ],
    run: { arguments in
        let orderId = arguments["orderId"] as? String ?? ""
        let status = OrderStore.shared.status(for: orderId)
        return ["orderId": orderId, "status": status]
    }
)

await server.register(lookup)
let session = try await makeSession(client: client)
let reply = try await session.respond(to: prompt)

The view still sends user text into respond. The model sees lookup_order because tools/list returned it. If you need a second tool tomorrow, add another MCPTool. Leave the view alone.

Checklist

1. Target iOS 27 or macOS 27, and Xcode 27.
2. Keep the MCP server free of Foundation Models imports.
3. Speak raw JSON-RPC for initialize, tools/list, and tools/call.
4. Map listed tools with a runtime schema, not a new compiled Tool per method.
5. Create or refresh the session after the list changes. Do not edit tools while isResponding is true.
6. Put file, network, and private-data tools through the sandbox and guardrail courses before you ship them.

Primary sources: MCP specification, Foundation Models, and WWDC26 session 242.

Related courses: Run tools in a safe box, Stop prompt attacks and leaks, and Core AI vs Core ML vs MLX.