diff --git a/src/mcp/bridge.ts b/src/mcp/bridge.ts
index ae9c6d6..01ab1d6 100644
--- a/src/mcp/bridge.ts
+++ b/src/mcp/bridge.ts
@@ -81,6 +81,12 @@ export function isCustomerOnline(customerId: string): boolean {
  * Forward a JSON-RPC MCP call to the customer's WebSocket and return the result.
  * Throws if customer is offline or times out.
  */
+// Monotonic counter to ensure each WS-forwarded call has a unique id.
+// Caller-provided ids (e.g. from LiteLLM) can collide across concurrent
+// sessions — LiteLLM opens a fresh session per tool call and each uses
+// id=1. We wrap with our own id on the wire and translate back.
+let nextWireId = 1;
+
 export async function forwardMcpCall(
   customerId: string,
   request: { method: string; params?: unknown; id: string | number },
@@ -91,19 +97,28 @@ export async function forwardMcpCall(
     throw new McpBridgeError("tool_unavailable: customer not online", -32001);
   }
 
+  const wireId = `demux-${nextWireId++}`;
+
   return new Promise((resolve, reject) => {
     const timer = setTimeout(() => {
-      session.pending.delete(request.id);
+      session.pending.delete(wireId);
       reject(new McpBridgeError(`tool_unavailable: timeout after ${timeoutMs}ms`, -32002));
     }, timeoutMs);
 
-    session.pending.set(request.id, { resolve, reject, timer });
+    session.pending.set(wireId, { resolve, reject, timer });
 
     try {
-      session.ws.send(JSON.stringify({ jsonrpc: "2.0", ...request }));
+      session.ws.send(
+        JSON.stringify({
+          jsonrpc: "2.0",
+          method: request.method,
+          params: request.params,
+          id: wireId,
+        })
+      );
     } catch (err) {
       clearTimeout(timer);
-      session.pending.delete(request.id);
+      session.pending.delete(wireId);
       reject(err);
     }
   });