diff --git a/src/instrumentation.ts b/src/instrumentation.ts
index 3b3ccd7d..f092019f 100644
--- a/src/instrumentation.ts
+++ b/src/instrumentation.ts
@@ -22,5 +22,53 @@ export async function register() {
     } catch (err) {
       console.error('Error reconciling sessions:', err);
     }
+
+    // Register graceful shutdown handler
+    registerShutdownHandler();
   }
 }
+
+/**
+ * Register signal handlers for graceful shutdown.
+ * Stops active Claude queries and disconnects Prisma so the process can exit cleanly.
+ */
+function registerShutdownHandler() {
+  let shuttingDown = false;
+
+  const shutdown = async (signal: string) => {
+    if (shuttingDown) {
+      console.log(`Received ${signal} again, forcing exit`);
+      process.exit(1);
+    }
+    shuttingDown = true;
+    console.log(`Received ${signal}, shutting down gracefully...`);
+
+    // Force exit after 10s if graceful shutdown hangs
+    // (important for SIGTERM from systemd where there's no second signal)
+    setTimeout(() => {
+      console.error('Graceful shutdown timed out, forcing exit');
+      process.exit(1);
+    }, 10_000).unref();
+
+    try {
+      // Stop all active Claude queries
+      const { stopAllSessions } = await import('@/server/services/claude-runner');
+      await stopAllSessions();
+    } catch (err) {
+      console.error('Error stopping sessions during shutdown:', err);
+    }
+
+    try {
+      // Disconnect Prisma
+      const { prisma } = await import('@/lib/prisma');
+      await prisma.$disconnect();
+    } catch (err) {
+      console.error('Error disconnecting Prisma during shutdown:', err);
+    }
+
+    process.exit(0);
+  };
+
+  process.on('SIGINT', () => void shutdown('SIGINT'));
+  process.on('SIGTERM', () => void shutdown('SIGTERM'));
+}
diff --git a/src/lib/rate-limiter.ts b/src/lib/rate-limiter.ts
index 09f81f8c..6b115c4c 100644
--- a/src/lib/rate-limiter.ts
+++ b/src/lib/rate-limiter.ts
@@ -212,11 +212,12 @@ export class RateLimiter {
 export const loginRateLimiter = new RateLimiter();
 
 // Start cleanup interval (every hour)
+// Use .unref() so this timer doesn't prevent graceful shutdown
 if (typeof setInterval !== 'undefined') {
   setInterval(
     () => {
       loginRateLimiter.cleanup();
     },
     60 * 60 * 1000
-  );
+  ).unref();
 }
diff --git a/src/server/services/claude-runner.ts b/src/server/services/claude-runner.ts
index 71b99183..0a4b4431 100644
--- a/src/server/services/claude-runner.ts
+++ b/src/server/services/claude-runner.ts
@@ -635,7 +635,7 @@ export async function markLastMessageAsInterrupted(sessionId: string): Promise<v
  * Stop a session's Claude query and clean up state.
  * Called when a session is stopped.
  */
-export async function stopSession(sessionId: string): Promise<void> {
+export function stopSession(sessionId: string): void {
   const state = sessions.get(sessionId);
   if (!state) return;
 
@@ -658,6 +658,17 @@ export async function stopSession(sessionId: string): Promise<void> {
   sessions.delete(sessionId);
 }
 
+/**
+ * Stop all active Claude queries. Called during graceful shutdown.
+ */
+export async function stopAllSessions(): Promise<void> {
+  const sessionIds = [...sessions.keys()];
+  if (sessionIds.length === 0) return;
+
+  log.info('Stopping all active sessions for shutdown', { count: sessionIds.length });
+  await Promise.allSettled(sessionIds.map((id) => stopSession(id)));
+}
+
 /**
  * Mark all running sessions as stopped.
  * Called on server startup since all in-memory state is lost.