diff --git a/core/http/endpoints/openai/mcp.go b/core/http/endpoints/openai/mcp.go
index fe018bbbd..a0e51f91e 100644
--- a/core/http/endpoints/openai/mcp.go
+++ b/core/http/endpoints/openai/mcp.go
@@ -72,12 +72,25 @@ func MCPCompletionEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader,
 			fragment = fragment.AddMessage(message.Role, message.StringContent)
 		}
 
-		port := appConfig.APIAddress[strings.LastIndex(appConfig.APIAddress, ":")+1:]
+		// Extract port from APIAddress (format: ":8080" or "127.0.0.1:8080" or "0.0.0.0:8080")
+		port := "8080" // default
+		if appConfig.APIAddress != "" {
+			lastColon := strings.LastIndex(appConfig.APIAddress, ":")
+			if lastColon >= 0 && lastColon+1 < len(appConfig.APIAddress) {
+				port = appConfig.APIAddress[lastColon+1:]
+			} else {
+				log.Warn().Str("APIAddress", appConfig.APIAddress).Msg("[MCP] Could not extract port from APIAddress, using default 8080")
+			}
+		}
+
 		apiKey := ""
-		if appConfig.ApiKeys != nil {
+		if len(appConfig.ApiKeys) > 0 {
 			apiKey = appConfig.ApiKeys[0]
 		}
 
+		baseURL := "http://127.0.0.1:" + port
+		log.Debug().Str("baseURL", baseURL).Str("model", config.Name).Msg("[MCP] Creating OpenAI LLM client for internal API calls")
+
 		ctxWithCancellation, cancel := context.WithCancel(ctx)
 		defer cancel()
 		handleConnectionCancellation(c, cancel, ctxWithCancellation)
@@ -85,7 +98,7 @@ func MCPCompletionEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader,
 		// and act like completion.go.
 		// We can do this as cogito expects an interface and we can create one that
 		// we satisfy to just call internally ComputeChoices
-		defaultLLM := cogito.NewOpenAILLM(config.Name, apiKey, "http://127.0.0.1:"+port)
+		defaultLLM := cogito.NewOpenAILLM(config.Name, apiKey, baseURL)
 
 		cogitoOpts := []cogito.Option{
 			cogito.WithStatusCallback(func(s string) {
@@ -127,7 +140,8 @@ func MCPCompletionEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader,
 			cogitoOpts...,
 		)
 		if err != nil && !errors.Is(err, cogito.ErrNoToolSelected) {
-			return err
+			log.Error().Err(err).Msgf("[MCP] ExecuteTools failed for model %s", config.Name)
+			return fmt.Errorf("failed to execute tools: %w", err)
 		}
 
 		f, err = defaultLLM.Ask(ctx, f)
diff --git a/core/http/static/chat.js b/core/http/static/chat.js
index 993c956ac..5bb1934d7 100644
--- a/core/http/static/chat.js
+++ b/core/http/static/chat.js
@@ -435,9 +435,28 @@ async function promptGPT(systemPrompt, input) {
   }
 
   if (!response.ok) {
+    // Try to get error details from response body
+    let errorMessage = `Error: POST ${endpoint} ${response.status}`;
+    try {
+      const errorData = await response.json();
+      if (errorData && errorData.error && errorData.error.message) {
+        errorMessage = `Error (${response.status}): ${errorData.error.message}`;
+      }
+    } catch (e) {
+      // If response is not JSON, try to get text
+      try {
+        const errorText = await response.text();
+        if (errorText) {
+          errorMessage = `Error (${response.status}): ${errorText.substring(0, 200)}`;
+        }
+      } catch (e2) {
+        // Ignore - use default error message
+      }
+    }
+    
     Alpine.store("chat").add(
       "assistant",
-      `<span class='error'>Error: POST ${endpoint} ${response.status}</span>`,
+      `<span class='error'>${errorMessage}</span>`,
     );
     toggleLoader(false);
     currentAbortController = null;