[responses API] Add list_tools_for_servers and threading server_keys in routers (#16540)

sgl-model-gateway/src/mcp/manager.rs

@@ -161,6 +161,39 @@ impl McpManager {
             .collect()
     }
 
+    /// List tools only from specific servers plus all static servers
+    ///
+    /// This method filters tools to only include:
+    /// 1. Tools from static servers (always visible)
+    /// 2. Tools from the specified dynamic servers
+    ///
+    /// This provides request-scoped tool isolation while maintaining
+    /// global visibility for static servers.
+    pub fn list_tools_for_servers(&self, server_keys: &[String]) -> Vec<Tool> {
+        self.inventory
+            .list_tools()
+            .into_iter()
+            .filter(|(_tool_name, server_key, _tool_info)| {
+                // Include if:
+                // 1. It's a static server (check by name in static_clients)
+                // 2. It's in the requested servers list
+                self.is_static_server_by_key(server_key) || server_keys.contains(server_key)
+            })
+            .map(|(_tool_name, _server_key, tool_info)| tool_info)
+            .collect()
+    }
+
+    /// Check if a server key belongs to a static server
+    ///
+    /// Static servers can be identified by checking if their name
+    /// exists in the static_clients map. We need to handle the fact
+    /// that static servers use name as key while dynamic use URL.
+    fn is_static_server_by_key(&self, server_key: &str) -> bool {
+        // For static servers, the server_key in inventory is the server name
+        // Check if this key exists in static_clients
+        self.static_clients.contains_key(server_key)
+    }
+
     /// Call a tool by name with automatic type coercion
     ///
     /// Accepts either JSON string or parsed Map as arguments.
@@ -736,7 +769,7 @@ impl McpManager {
     }
 
     /// Generate a unique key for a server config
-    fn server_key(config: &McpServerConfig) -> String {
+    pub fn server_key(config: &McpServerConfig) -> String {
         // Extract URL from transport or use name
         match &config.transport {
             McpTransport::Streamable { url, .. } => url.clone(),

sgl-model-gateway/src/routers/grpc/common/responses/utils.rs

@@ -22,11 +22,12 @@ use crate::{
 /// Ensure MCP connection succeeds if MCP tools are declared
 ///
 /// Checks if request declares MCP tools, and if so, validates that
-/// the MCP client can be created and connected.
+/// the MCP clients can be created and connected.
+/// Returns Ok((has_mcp_tools, server_keys)) on success.
 pub(crate) async fn ensure_mcp_connection(
     mcp_manager: &Arc<McpManager>,
     tools: Option<&[ResponseTool]>,
-) -> Result<bool, Response> {
+) -> Result<(bool, Vec<String>), Response> {
     let has_mcp_tools = tools
         .map(|t| {
             t.iter()
@@ -36,23 +37,25 @@ pub(crate) async fn ensure_mcp_connection(
 
     if has_mcp_tools {
         if let Some(tools) = tools {
-            if ensure_request_mcp_client(mcp_manager, tools)
+            match ensure_request_mcp_client(mcp_manager, tools).await {
-                .await
+                Some((_manager, server_keys)) => {
-                .is_none()
+                    return Ok((true, server_keys));
-            {
+                }
+                None => {
                     error!(
                         function = "ensure_mcp_connection",
-                    "Failed to connect to MCP server"
+                        "Failed to connect to MCP servers"
                     );
                     return Err(error::failed_dependency(
                         "connect_mcp_server_failed",
-                    "Failed to connect to MCP server. Check server_url and authorization.",
+                        "Failed to connect to MCP servers. Check server_url and authorization.",
                     ));
                 }
             }
         }
+    }
 
-    Ok(has_mcp_tools)
+    Ok((false, Vec::new()))
 }
 
 /// Validate that workers are available for the requested model

sgl-model-gateway/src/routers/grpc/harmony/responses/common.rs

@@ -1,7 +1,5 @@
 //! Shared helpers and state tracking for Harmony Responses
 
-use std::sync::Arc;
-
 use axum::response::Response;
 use serde_json::{from_value, json, to_string, Value};
 use tracing::{debug, error, warn};
@@ -10,7 +8,7 @@ use uuid::Uuid;
 use super::{context::HarmonyResponsesContext, execution::ToolResult};
 use crate::{
     data_connector::ResponseId,
-    mcp::McpManager,
+    mcp,
     protocols::{
         common::{ToolCall, ToolChoice, ToolChoiceValue},
         responses::{
@@ -217,10 +215,10 @@ pub(super) fn build_next_request_with_tools(
 pub(super) fn inject_mcp_metadata(
     response: &mut ResponsesResponse,
     tracking: &McpCallTracking,
-    mcp_manager: &Arc<McpManager>,
+    mcp_tools: &[mcp::Tool],
 ) {
     // Build mcp_list_tools item
-    let tools = mcp_manager.list_tools();
+    let tools = mcp_tools;
     let tools_info: Vec<McpToolInfo> = tools
         .iter()
         .map(|t| McpToolInfo {

sgl-model-gateway/src/routers/grpc/harmony/responses/context.rs

@@ -1,6 +1,6 @@
 //! Context for Harmony Responses execution
 
-use std::sync::Arc;
+use std::sync::{Arc, RwLock as StdRwLock};
 
 use crate::{
     data_connector::{ConversationItemStorage, ConversationStorage, ResponseStorage},
@@ -23,6 +23,9 @@ pub(crate) struct HarmonyResponsesContext {
     /// MCP manager for tool execution
     pub mcp_manager: Arc<McpManager>,
 
+    /// Server keys for MCP tools requested in this context
+    pub requested_servers: Arc<StdRwLock<Vec<String>>>,
+
     /// Response storage for loading conversation history
     pub response_storage: Arc<dyn ResponseStorage>,
 
@@ -47,6 +50,7 @@ impl HarmonyResponsesContext {
             pipeline,
             components,
             mcp_manager,
+            requested_servers: Arc::new(StdRwLock::new(Vec::new())),
             response_storage,
             conversation_storage,
             conversation_item_storage,

sgl-model-gateway/src/routers/grpc/harmony/responses/non_streaming.rs

@@ -57,9 +57,15 @@ pub(crate) async fn serve_harmony_responses(
     let current_request = load_previous_messages(ctx, request).await?;
 
     // Check MCP connection and get whether MCP tools are present
-    let has_mcp_tools =
+    let (has_mcp_tools, server_keys) =
         ensure_mcp_connection(&ctx.mcp_manager, current_request.tools.as_deref()).await?;
 
+    // Set the server keys in the context
+    {
+        let mut servers = ctx.requested_servers.write().unwrap();
+        *servers = server_keys;
+    }
+
     let response = if has_mcp_tools {
         execute_with_mcp_loop(ctx, current_request).await?
     } else {
@@ -96,8 +102,11 @@ async fn execute_with_mcp_loop(
     // Extract user's max_tool_calls limit (if set)
     let max_tool_calls = current_request.max_tool_calls.map(|n| n as usize);
 
-    // Add static MCP tools from inventory to the request
+    // Add filtered MCP tools (static + requested dynamic) to the request
-    let mcp_tools = ctx.mcp_manager.list_tools();
+    let mcp_tools = {
+        let servers = ctx.requested_servers.read().unwrap();
+        ctx.mcp_manager.list_tools_for_servers(&servers)
+    };
     if !mcp_tools.is_empty() {
         let mcp_response_tools = convert_mcp_tools_to_response_tools(&mcp_tools);
 
@@ -216,7 +225,7 @@ async fn execute_with_mcp_loop(
 
                     // Inject MCP metadata if any calls were executed
                     if mcp_tracking.total_calls() > 0 {
-                        inject_mcp_metadata(&mut response, &mcp_tracking, &ctx.mcp_manager);
+                        inject_mcp_metadata(&mut response, &mcp_tracking, &mcp_tools);
                     }
 
                     return Ok(response);
@@ -258,7 +267,7 @@ async fn execute_with_mcp_loop(
 
                     // Inject MCP metadata for all executed calls
                     if mcp_tracking.total_calls() > 0 {
-                        inject_mcp_metadata(&mut response, &mcp_tracking, &ctx.mcp_manager);
+                        inject_mcp_metadata(&mut response, &mcp_tracking, &mcp_tools);
                     }
 
                     return Ok(response);
@@ -291,7 +300,7 @@ async fn execute_with_mcp_loop(
                 );
 
                 // Inject MCP metadata into final response
-                inject_mcp_metadata(&mut response, &mcp_tracking, &ctx.mcp_manager);
+                inject_mcp_metadata(&mut response, &mcp_tracking, &mcp_tools);
 
                 debug!(
                     mcp_calls = mcp_tracking.total_calls(),

sgl-model-gateway/src/routers/grpc/harmony/responses/streaming.rs

@@ -47,12 +47,18 @@ pub(crate) async fn serve_harmony_responses_stream(
     };
 
     // Check MCP connection BEFORE starting stream and get whether MCP tools are present
-    let has_mcp_tools =
+    let (has_mcp_tools, server_keys) =
         match ensure_mcp_connection(&ctx.mcp_manager, current_request.tools.as_deref()).await {
-            Ok(has_mcp) => has_mcp,
+            Ok(result) => result,
             Err(response) => return response,
         };
 
+    // Set the server keys in the context
+    {
+        let mut servers = ctx.requested_servers.write().unwrap();
+        *servers = server_keys;
+    }
+
     // Create SSE channel
     let (tx, rx) = mpsc::unbounded_channel();
 
@@ -124,8 +130,11 @@ async fn execute_mcp_tool_loop_streaming(
     // Extract user's max_tool_calls limit (if set)
     let max_tool_calls = current_request.max_tool_calls.map(|n| n as usize);
 
-    // Add static MCP tools from inventory
+    // Add filtered MCP tools (static + requested dynamic) to the request
-    let mcp_tools = ctx.mcp_manager.list_tools();
+    let mcp_tools = {
+        let servers = ctx.requested_servers.read().unwrap();
+        ctx.mcp_manager.list_tools_for_servers(&servers)
+    };
     if !mcp_tools.is_empty() {
         let mcp_response_tools = convert_mcp_tools_to_response_tools(&mcp_tools);
         let mut all_tools = current_request.tools.clone().unwrap_or_default();

sgl-model-gateway/src/routers/grpc/regular/responses/common.rs

@@ -175,8 +175,9 @@ pub(super) fn generate_mcp_id(prefix: &str) -> String {
 pub(super) fn build_mcp_list_tools_item(
     mcp: &Arc<McpManager>,
     server_label: &str,
+    server_keys: &[String],
 ) -> ResponseOutputItem {
-    let tools = mcp.list_tools();
+    let tools = mcp.list_tools_for_servers(server_keys);
     let tools_info: Vec<McpToolInfo> = tools
         .iter()
         .map(|t| McpToolInfo {

sgl-model-gateway/src/routers/grpc/regular/responses/context.rs

@@ -3,7 +3,10 @@
 //! Bundles all dependencies needed by responses handlers to avoid passing
 //! 10+ parameters to every function.
 
-use std::{collections::HashMap, sync::Arc};
+use std::{
+    collections::HashMap,
+    sync::{Arc, RwLock as StdRwLock},
+};
 
 use tokio::{sync::RwLock, task::JoinHandle};
 
@@ -50,6 +53,9 @@ pub(crate) struct ResponsesContext {
     /// MCP manager for tool support
     pub mcp_manager: Arc<McpManager>,
 
+    /// Server keys for MCP tools requested in this context
+    pub requested_servers: Arc<StdRwLock<Vec<String>>>,
+
     /// Background task handles for cancellation support
     pub background_tasks: Arc<RwLock<HashMap<String, BackgroundTaskInfo>>>,
 }
@@ -71,6 +77,7 @@ impl ResponsesContext {
             conversation_storage,
             conversation_item_storage,
             mcp_manager,
+            requested_servers: Arc::new(StdRwLock::new(Vec::new())),
             background_tasks: Arc::new(RwLock::new(HashMap::new())),
         }
     }

sgl-model-gateway/src/routers/grpc/regular/responses/handlers.rs

@@ -110,12 +110,18 @@ async fn route_responses_streaming(
     };
 
     // 2. Check MCP connection and get whether MCP tools are present
-    let has_mcp_tools =
+    let (has_mcp_tools, server_keys) =
         match ensure_mcp_connection(&ctx.mcp_manager, request.tools.as_deref()).await {
-            Ok(has_mcp) => has_mcp,
+            Ok(result) => result,
             Err(response) => return response,
         };
 
+    // Set the server keys in the context
+    {
+        let mut servers = ctx.requested_servers.write().unwrap();
+        *servers = server_keys;
+    }
+
     if has_mcp_tools {
         debug!("MCP tools detected in streaming mode, using streaming tool loop");
 

sgl-model-gateway/src/routers/grpc/regular/responses/non_streaming.rs

@@ -48,7 +48,14 @@ pub(super) async fn route_responses_internal(
     let modified_request = load_conversation_history(ctx, &request).await?;
 
     // 2. Check MCP connection and get whether MCP tools are present
-    let has_mcp_tools = ensure_mcp_connection(&ctx.mcp_manager, request.tools.as_deref()).await?;
+    let (has_mcp_tools, server_keys) =
+        ensure_mcp_connection(&ctx.mcp_manager, request.tools.as_deref()).await?;
+
+    // Set the server keys in the context
+    {
+        let mut servers = ctx.requested_servers.write().unwrap();
+        *servers = server_keys;
+    }
 
     let responses_response = if has_mcp_tools {
         debug!("MCP tools detected, using tool loop");
@@ -167,7 +174,10 @@ pub(super) async fn execute_tool_loop(
     );
 
     // Get MCP tools and convert to chat format (do this once before loop)
-    let mcp_tools = ctx.mcp_manager.list_tools();
+    let mcp_tools = {
+        let servers = ctx.requested_servers.read().unwrap();
+        ctx.mcp_manager.list_tools_for_servers(&servers)
+    };
     let mcp_chat_tools = convert_mcp_tools_to_chat_tools(&mcp_tools);
     trace!(
         "Converted {} MCP tools to chat format",
@@ -399,7 +409,9 @@ pub(super) async fn execute_tool_loop(
             // Inject MCP metadata into output
             if state.total_calls > 0 {
                 // Prepend mcp_list_tools item
-                let mcp_list_tools = build_mcp_list_tools_item(&ctx.mcp_manager, &server_label);
+                let servers = ctx.requested_servers.read().unwrap();
+                let mcp_list_tools =
+                    build_mcp_list_tools_item(&ctx.mcp_manager, &server_label, &servers);
                 responses_response.output.insert(0, mcp_list_tools);
 
                 // Append all mcp_call items at the end

sgl-model-gateway/src/routers/mcp_utils.rs

@@ -34,12 +34,16 @@ pub struct McpLoopConfig {
     /// Maximum iterations as safety limit (default: DEFAULT_MAX_ITERATIONS).
     /// Prevents infinite loops when max_tool_calls is not set by user.
     pub max_iterations: usize,
+    /// Server keys for filtering MCP tools.
+    /// Contains keys for dynamic servers that were connected for this request.
+    pub server_keys: Vec<String>,
 }
 
 impl Default for McpLoopConfig {
     fn default() -> Self {
         Self {
             max_iterations: DEFAULT_MAX_ITERATIONS,
+            server_keys: Vec::new(),
         }
     }
 }
@@ -70,23 +74,27 @@ pub fn extract_server_label(tools: Option<&[ResponseTool]>, default_label: &str)
 // MCP Connection
 // ============================================================================
 
-/// Ensure MCP client is connected for request-level MCP tools.
+/// Ensure MCP clients are connected for all request-level MCP tools.
 ///
-/// This function extracts MCP server configuration from request tools (server_url, authorization)
+/// This function extracts MCP server configurations from ALL request tools (server_url, authorization)
-/// and ensures a client connection is established via the connection pool.
+/// and ensures client connections are established via the connection pool.
 ///
-/// Returns `Some(())` if a dynamic MCP tool was found and client was created/retrieved,
+/// Returns `Some((manager, server_keys))` if MCP tools were found and clients created,
-/// `None` if no MCP tools with server_url were found or connection failed.
+/// `None` if no MCP tools with server_url were found.
 pub async fn ensure_request_mcp_client(
     mcp_manager: &Arc<McpManager>,
     tools: &[ResponseTool],
-) -> Option<()> {
+) -> Option<(Arc<McpManager>, Vec<String>)> {
-    // Find an MCP tool with a server_url
+    let mut server_keys = Vec::new();
-    let tool = tools
+    let mut has_mcp_tools = false;
-        .iter()
-        .find(|t| matches!(t.r#type, ResponseToolType::Mcp) && t.server_url.is_some())?;
 
-    let server_url = tool.server_url.as_ref()?.trim().to_string();
+    // Process all MCP tools
+    for tool in tools {
+        if matches!(tool.r#type, ResponseToolType::Mcp) && tool.server_url.is_some() {
+            has_mcp_tools = true;
+            let Some(server_url) = tool.server_url.as_ref().map(|s| s.trim().to_string()) else {
+                continue;
+            };
 
             // Validate URL scheme
             if !(server_url.starts_with("http://") || server_url.starts_with("https://")) {
@@ -94,7 +102,7 @@ pub async fn ensure_request_mcp_client(
                     "Ignoring MCP server_url with unsupported scheme: {}",
                     server_url
                 );
-        return None;
+                continue;
             }
 
             // Extract server label and auth token
@@ -125,12 +133,31 @@ pub async fn ensure_request_mcp_client(
                 required: false,
             };
 
+            // Get the server key for tracking
+            let server_key = McpManager::server_key(&server_config);
+
             // Use get_or_create_client to establish connection
             match mcp_manager.get_or_create_client(server_config).await {
-        Ok(_client) => Some(()),
+                Ok(_client) => {
+                    // Track this server for filtering
+                    if !server_keys.contains(&server_key) {
+                        server_keys.push(server_key);
+                    }
+                }
                 Err(err) => {
-            warn!("Failed to get/create MCP connection: {}", err);
+                    warn!(
+                        "Failed to get/create MCP connection for {}: {}",
+                        server_key, err
+                    );
+                    // Continue processing other tools
+                }
+            }
+        }
+    }
+
+    if has_mcp_tools && !server_keys.is_empty() {
+        Some((mcp_manager.clone(), server_keys))
+    } else {
         None
     }
 }
-}

sgl-model-gateway/src/routers/openai/context.rs

@@ -238,6 +238,7 @@ pub struct StreamingEventContext<'a> {
     pub server_label: &'a str,
     pub original_request: &'a ResponsesRequest,
     pub previous_response_id: Option<&'a str>,
+    pub server_keys: &'a [String],
 }
 
 pub type StreamingRequest = OwnedStreamingContext;

sgl-model-gateway/src/routers/openai/responses/mcp.rs

@@ -202,6 +202,7 @@ pub(super) async fn execute_streaming_tool_calls(
 pub(super) fn prepare_mcp_payload_for_streaming(
     payload: &mut Value,
     active_mcp: &Arc<mcp::McpManager>,
+    server_keys: &[String],
 ) {
     if let Some(obj) = payload.as_object_mut() {
         // Remove any non-function tools from outgoing payload
@@ -217,7 +218,7 @@ pub(super) fn prepare_mcp_payload_for_streaming(
         }
 
         // Build function tools for all discovered MCP tools
-        let tools = active_mcp.list_tools();
+        let tools = active_mcp.list_tools_for_servers(server_keys);
         let mut tools_json = Vec::with_capacity(tools.len());
         for t in tools {
             let parameters = Value::Object((*t.input_schema).clone());
@@ -310,8 +311,9 @@ pub(super) fn send_mcp_list_tools_events(
     server_label: &str,
     output_index: usize,
     sequence_number: &mut u64,
+    server_keys: &[String],
 ) -> bool {
-    let tools_item_full = build_mcp_list_tools_item(mcp, server_label);
+    let tools_item_full = build_mcp_list_tools_item(mcp, server_label, server_keys);
     let item_id = tools_item_full
         .get("id")
         .and_then(|v| v.as_str())
@@ -464,13 +466,14 @@ pub(super) fn inject_mcp_metadata_streaming(
     state: &ToolLoopState,
     mcp: &Arc<mcp::McpManager>,
     server_label: &str,
+    server_keys: &[String],
 ) {
     if let Some(output_array) = response.get_mut("output").and_then(|v| v.as_array_mut()) {
         output_array.retain(|item| {
             item.get("type").and_then(|t| t.as_str()) != Some(ItemType::MCP_LIST_TOOLS)
         });
 
-        let list_tools_item = build_mcp_list_tools_item(mcp, server_label);
+        let list_tools_item = build_mcp_list_tools_item(mcp, server_label, server_keys);
         output_array.insert(0, list_tools_item);
 
         let mcp_call_items =
@@ -482,7 +485,7 @@ pub(super) fn inject_mcp_metadata_streaming(
         }
     } else if let Some(obj) = response.as_object_mut() {
         let mut output_items = Vec::new();
-        output_items.push(build_mcp_list_tools_item(mcp, server_label));
+        output_items.push(build_mcp_list_tools_item(mcp, server_label, server_keys));
         output_items.extend(build_executed_mcp_call_items(
             &state.conversation_history,
             server_label,
@@ -584,6 +587,7 @@ pub(super) async fn execute_tool_loop(
                     "max_tool_calls",
                     active_mcp,
                     original_body,
+                    &config.server_keys,
                 );
             }
 
@@ -634,7 +638,8 @@ pub(super) async fn execute_tool_loop(
                 let server_label = extract_server_label(original_body.tools.as_deref(), "mcp");
 
                 // Build mcp_list_tools item
-                let list_tools_item = build_mcp_list_tools_item(active_mcp, &server_label);
+                let list_tools_item =
+                    build_mcp_list_tools_item(active_mcp, &server_label, &config.server_keys);
 
                 // Insert at beginning of output array
                 if let Some(output_array) = response_json
@@ -668,6 +673,7 @@ pub(super) fn build_incomplete_response(
     reason: &str,
     active_mcp: &Arc<mcp::McpManager>,
     original_body: &ResponsesRequest,
+    server_keys: &[String],
 ) -> Result<Value, String> {
     let obj = response
         .as_object_mut()
@@ -712,7 +718,7 @@ pub(super) fn build_incomplete_response(
 
         // Add mcp_list_tools and executed mcp_call items at the beginning
         if state.total_calls > 0 || !mcp_call_items.is_empty() {
-            let list_tools_item = build_mcp_list_tools_item(active_mcp, &server_label);
+            let list_tools_item = build_mcp_list_tools_item(active_mcp, &server_label, server_keys);
             output_array.insert(0, list_tools_item);
 
             // Add mcp_call items for executed calls using helper
@@ -758,8 +764,12 @@ pub(super) fn build_incomplete_response(
 // ============================================================================
 
 /// Build a mcp_list_tools output item
-pub(super) fn build_mcp_list_tools_item(mcp: &Arc<mcp::McpManager>, server_label: &str) -> Value {
+pub(super) fn build_mcp_list_tools_item(
-    let tools = mcp.list_tools();
+    mcp: &Arc<mcp::McpManager>,
+    server_label: &str,
+    server_keys: &[String],
+) -> Value {
+    let tools = mcp.list_tools_for_servers(server_keys);
     let tools_json: Vec<Value> = tools
         .iter()
         .map(|t| {

sgl-model-gateway/src/routers/openai/responses/non_streaming.rs

@@ -50,11 +50,15 @@ pub async fn handle_non_streaming_response(mut ctx: RequestContext) -> Response
         }
     };
 
-    if let Some(ref tools) = original_body.tools {
+    let server_keys = match original_body.tools.as_ref() {
-        ensure_request_mcp_client(mcp_manager, tools.as_slice()).await;
+        Some(tools) => match ensure_request_mcp_client(mcp_manager, tools.as_slice()).await {
-    }
+            Some((_manager, keys)) => keys,
+            None => Vec::new(),
+        },
+        None => Vec::new(),
+    };
 
-    let active_mcp = if mcp_manager.list_tools().is_empty() {
+    let active_mcp = if mcp_manager.list_tools_for_servers(&server_keys).is_empty() {
         None
     } else {
         Some(mcp_manager)
@@ -63,8 +67,11 @@ pub async fn handle_non_streaming_response(mut ctx: RequestContext) -> Response
     let mut response_json: Value;
 
     if let Some(mcp) = active_mcp {
-        let config = McpLoopConfig::default();
+        let config = McpLoopConfig {
-        prepare_mcp_payload_for_streaming(&mut payload, mcp);
+            server_keys: server_keys.clone(),
+            ..McpLoopConfig::default()
+        };
+        prepare_mcp_payload_for_streaming(&mut payload, mcp, &server_keys);
 
         match execute_tool_loop(
             ctx.components.client(),

sgl-model-gateway/src/routers/openai/responses/streaming.rs

@@ -442,7 +442,13 @@ pub(super) fn send_final_response_event(
     }
 
     if let Some(mcp) = active_mcp {
-        inject_mcp_metadata_streaming(&mut final_response, state, mcp, ctx.server_label);
+        inject_mcp_metadata_streaming(
+            &mut final_response,
+            state,
+            mcp,
+            ctx.server_label,
+            ctx.server_keys,
+        );
     }
 
     mask_tools_as_mcp(&mut final_response, ctx.original_request);
@@ -632,10 +638,11 @@ pub(super) async fn handle_streaming_with_tool_interception(
     headers: Option<&HeaderMap>,
     req: StreamingRequest,
     active_mcp: &Arc<crate::mcp::McpManager>,
+    server_keys: Vec<String>,
 ) -> Response {
     // Transform MCP tools to function tools in payload
     let mut payload = req.payload;
-    prepare_mcp_payload_for_streaming(&mut payload, active_mcp);
+    prepare_mcp_payload_for_streaming(&mut payload, active_mcp, &server_keys);
 
     let (tx, rx) = mpsc::unbounded_channel::<Result<Bytes, io::Error>>();
     let should_store = req.original_body.store.unwrap_or(false);
@@ -650,11 +657,15 @@ pub(super) async fn handle_streaming_with_tool_interception(
     let headers_opt = headers.cloned();
     let payload_clone = payload.clone();
     let active_mcp_clone = Arc::clone(active_mcp);
+    let server_keys_clone = server_keys.clone();
 
     // Spawn the streaming loop task
     tokio::spawn(async move {
         let mut state = ToolLoopState::new(original_request.input.clone());
-        let loop_config = McpLoopConfig::default();
+        let loop_config = McpLoopConfig {
+            server_keys: server_keys_clone.clone(),
+            ..McpLoopConfig::default()
+        };
         let max_tool_calls = original_request.max_tool_calls.map(|n| n as usize);
         let tools_json = payload_clone.get("tools").cloned().unwrap_or(json!([]));
         let base_payload = payload_clone.clone();
@@ -680,6 +691,7 @@ pub(super) async fn handle_streaming_with_tool_interception(
             server_label,
             original_request: &original_request,
             previous_response_id: previous_response_id.as_deref(),
+            server_keys: &server_keys_clone,
         };
 
         loop {
@@ -789,6 +801,7 @@ pub(super) async fn handle_streaming_with_tool_interception(
                                                         server_label,
                                                         list_tools_index,
                                                         &mut sequence_number,
+                                                        &server_keys_clone,
                                                     ) {
                                                         // Client disconnected
                                                         return;
@@ -868,6 +881,7 @@ pub(super) async fn handle_streaming_with_tool_interception(
                         &state,
                         &active_mcp_clone,
                         server_label,
+                        &server_keys_clone,
                     );
 
                     mask_tools_as_mcp(&mut response_json, &original_request);
@@ -977,11 +991,15 @@ pub async fn handle_streaming_response(ctx: RequestContext) -> Response {
     let original_body = ctx.responses_request();
     let mcp_manager = ctx.components.mcp_manager().expect("MCP manager required");
 
-    if let Some(ref tools) = original_body.tools {
+    let server_keys = match original_body.tools.as_ref() {
-        ensure_request_mcp_client(mcp_manager, tools.as_slice()).await;
+        Some(tools) => match ensure_request_mcp_client(mcp_manager, tools.as_slice()).await {
-    }
+            Some((_manager, keys)) => keys,
+            None => Vec::new(),
+        },
+        None => Vec::new(),
+    };
 
-    let active_mcp = if mcp_manager.list_tools().is_empty() {
+    let active_mcp = if mcp_manager.list_tools_for_servers(&server_keys).is_empty() {
         None
     } else {
         Some(mcp_manager.clone())
@@ -1003,5 +1021,12 @@ pub async fn handle_streaming_response(ctx: RequestContext) -> Response {
     let active_mcp = active_mcp.unwrap();
 
     // MCP is active - transform tools and set up interception
-    handle_streaming_with_tool_interception(&client, headers.as_ref(), req, &active_mcp).await
+    handle_streaming_with_tool_interception(
+        &client,
+        headers.as_ref(),
+        req,
+        &active_mcp,
+        server_keys,
+    )
+    .await
 }