diff --git a/clawshell.example.toml b/clawshell.example.toml
index fa0eb11..2fb92f5 100644
--- a/clawshell.example.toml
+++ b/clawshell.example.toml
@@ -7,6 +7,10 @@ log_level = "info"
 host = "127.0.0.1"
 port = 18790
 
+# Graceful shutdown timeout (seconds). When receiving SIGTERM/SIGINT,
+# the server waits up to this duration for in-flight requests to complete.
+# shutdown_timeout_secs = 30
+
 [upstream]
 base_url = "https://api.openai.com"
 anthropic_base_url = "https://api.anthropic.com"
diff --git a/src/config.rs b/src/config.rs
index 9c10138..8769b74 100644
--- a/src/config.rs
+++ b/src/config.rs
@@ -42,6 +42,13 @@ pub struct ServerConfig {
     pub host: String,
     #[serde(default = "default_port")]
     pub port: u16,
+
+    /// Graceful shutdown timeout in seconds. When receiving a shutdown signal
+    /// (SIGTERM/SIGINT), the server will wait up to this duration for in-flight
+    /// requests to complete before forcefully terminating.
+    /// Default: 30 seconds.
+    #[serde(default = "default_shutdown_timeout_secs")]
+    pub shutdown_timeout_secs: u64,
 }
 
 fn default_host() -> String {
@@ -52,6 +59,10 @@ fn default_port() -> u16 {
     18790
 }
 
+fn default_shutdown_timeout_secs() -> u64 {
+    30
+}
+
 #[derive(Debug, Deserialize, Clone)]
 pub struct UpstreamConfig {
     #[serde(default = "default_base_url")]
diff --git a/src/main.rs b/src/main.rs
index 0a452a3..b38ab8a 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -131,21 +131,50 @@ async fn cmd_start_inner(
 
     process::drop_privileges()?;
 
-    axum::serve(listener, app)
-        .with_graceful_shutdown(async {
-            let ctrl_c = signal::ctrl_c();
-            #[cfg(unix)]
-            let mut term = signal::unix::signal(signal::unix::SignalKind::terminate()).unwrap();
-            #[cfg(unix)]
-            tokio::select! { _ = ctrl_c => {}, _ = term.recv() => {} };
-            #[cfg(not(unix))]
+    // Capture shutdown timeout from config before moving into async block
+    let shutdown_timeout_secs = config.server.shutdown_timeout_secs;
+
+    // Create a shutdown signal future that handles both SIGTERM and SIGINT
+    let shutdown_signal = async move {
+        let ctrl_c = signal::ctrl_c();
+
+        #[cfg(unix)]
+        {
+            let mut term =
+                signal::unix::signal(signal::unix::SignalKind::terminate()).unwrap();
+            tokio::select! {
+                _ = ctrl_c => {
+                    info!("Received SIGINT (Ctrl+C), initiating graceful shutdown...");
+                }
+                _ = term.recv() => {
+                    info!("Received SIGTERM, initiating graceful shutdown...");
+                }
+            };
+        }
+
+        #[cfg(not(unix))]
+        {
             ctrl_c.await.ok();
-            info!("Shutdown signal received");
-        })
+            info!("Received shutdown signal, initiating graceful shutdown...");
+        }
+
+        info!(
+            timeout_secs = shutdown_timeout_secs,
+            "Waiting for in-flight requests to complete..."
+        );
+    };
+
+    // Start server with graceful shutdown
+    // Note: axum::serve's with_graceful_shutdown will:
+    // 1. Stop accepting new connections when shutdown signal is received
+    // 2. Wait for existing connections to complete their requests
+    // 3. The timeout is applied at the TCP level by the OS
+    axum::serve(listener, app)
+        .with_graceful_shutdown(shutdown_signal)
         .await?;
 
     process::remove_pid_file();
-    info!("ClawShell shut down");
+    info!("ClawShell shut down gracefully");
     Ok(())
 }