diff --git a/Cargo.lock b/Cargo.lock
index 24bc347..767d617 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -2032,6 +2032,7 @@ dependencies = [
  "anyhow",
  "clap",
  "image",
+ "libc",
  "obscura-browser",
  "obscura-cdp",
  "obscura-dom",
diff --git a/crates/obscura-cli/Cargo.toml b/crates/obscura-cli/Cargo.toml
index dc99798..bebc740 100644
--- a/crates/obscura-cli/Cargo.toml
+++ b/crates/obscura-cli/Cargo.toml
@@ -34,5 +34,8 @@ url = { workspace = true }
 serde = { workspace = true }
 serde_json = { workspace = true }
 
+[target.'cfg(unix)'.dependencies]
+libc = "0.2"
+
 [dev-dependencies]
 image = { version = "0.25", default-features = false, features = ["png"] }
diff --git a/crates/obscura-cli/src/main.rs b/crates/obscura-cli/src/main.rs
index fd19f24..e0522bc 100644
--- a/crates/obscura-cli/src/main.rs
+++ b/crates/obscura-cli/src/main.rs
@@ -83,7 +83,8 @@ enum Command {
         /// Maximum live CDP connections. Each connection runs on its own OS
         /// thread with its own V8 isolates, so this bounds the server's thread
         /// and memory footprint. Connections beyond the limit are refused with
-        /// a 503 rather than queued.
+        /// a 503 rather than queued. With --workers N the limit applies to each
+        /// worker process, so the server-wide maximum is N times this value.
         #[arg(long, default_value_t = obscura_cdp::DEFAULT_MAX_CONNECTIONS)]
         max_connections: usize,
 
@@ -473,6 +474,7 @@ async fn run_cli() -> anyhow::Result<()> {
                     stealth,
                     user_agent,
                     font_dirs,
+                    max_connections,
                 )
                 .await?;
             } else {
@@ -637,12 +639,17 @@ async fn run_multi_worker_serve(
     stealth: bool,
     user_agent: Option<String>,
     font_dirs: Vec<std::path::PathBuf>,
+    max_connections: usize,
 ) -> anyhow::Result<()> {
     use tokio::io::AsyncWriteExt as _;
     use tokio::net::TcpListener;
     use std::sync::atomic::{AtomicBool, Ordering};
 
     let exe = std::env::current_exe()?;
+    // Register before spawning workers so an early SIGTERM is not fatal to the
+    // balancer alone and leaves no orphans.
+    let shutdown = shutdown_signal()?;
+    tokio::pin!(shutdown);
     // Claim the public port before starting children so another process cannot
     // take it during worker startup.
     let listener = TcpListener::bind((host.as_str(), port)).await?;
@@ -663,6 +670,9 @@ async fn run_multi_worker_serve(
         let mut cmd = TokioCommand::new(&exe);
         cmd.kill_on_drop(true);
         cmd.arg("serve").arg("--port").arg(worker_port.to_string());
+        // Same limit per worker, not split: connections are not spread evenly,
+        // so a split would refuse clients while other workers have room.
+        cmd.arg("--max-connections").arg(max_connections.to_string());
         // Workers receive the client-facing Host header through the TCP
         // load balancer. Let their CDP security gate accept that public port
         // while it continues to reject foreign hosts and browser origins.
@@ -703,18 +713,29 @@ async fn run_multi_worker_serve(
 
     let mut availability = Vec::with_capacity(workers as usize);
     let mut supervisors = tokio::task::JoinSet::new();
+    let (stop_tx, stop_rx) = tokio::sync::watch::channel(());
     for ((mut child, mut cmd), &worker_port) in children.into_iter().zip(&worker_ports) {
         let ready = Arc::new(AtomicBool::new(true));
         availability.push(ready.clone());
+        let mut stop = stop_rx.clone();
         supervisors.spawn(async move {
             loop {
-                let status = child.wait().await;
+                let status = tokio::select! {
+                    status = child.wait() => status,
+                    _ = stop.changed() => {
+                        stop_worker(&mut child).await;
+                        return;
+                    }
+                };
                 ready.store(false, Ordering::Relaxed);
                 tracing::warn!("worker on port {} exited: {:?}", worker_port, status);
                 loop {
                     // ponytail: fixed retry bounds crash loops; add backoff if
                     // persistently failing worker configurations need it.
                     tokio::time::sleep(Duration::from_millis(100)).await;
+                    if stop.has_changed().unwrap_or(true) {
+                        return;
+                    }
                     match cmd.spawn() {
                         Ok(mut replacement) => {
                             let deadline = tokio::time::Instant::now() + Duration::from_secs(5);
@@ -744,7 +765,10 @@ async fn run_multi_worker_serve(
     let mut next_worker = 0usize;
 
     loop {
-        let (client_stream, peer_addr) = listener.accept().await?;
+        let (client_stream, peer_addr) = tokio::select! {
+            accepted = listener.accept() => accepted?,
+            _ = &mut shutdown => break,
+        };
         if let Err(error) = client_stream.set_nodelay(true) {
             tracing::warn!("client {} TCP_NODELAY failed: {}", peer_addr, error);
         }
@@ -858,6 +882,56 @@ async fn run_multi_worker_serve(
             }
         });
     }
+
+    tracing::info!("Shutting down, stopping workers");
+    drop(stop_tx);
+    // Workers are stopped in parallel by their supervisors; anything left after
+    // the grace period is aborted, and kill_on_drop then kills the child.
+    let _ = tokio::time::timeout(Duration::from_secs(5), async {
+        while supervisors.join_next().await.is_some() {}
+    })
+    .await;
+    Ok(())
+}
+
+/// Resolves on SIGTERM or SIGINT (Ctrl-C elsewhere). Handlers are installed
+/// when this is called, not on first poll.
+fn shutdown_signal() -> std::io::Result<impl std::future::Future<Output = ()>> {
+    #[cfg(unix)]
+    {
+        use tokio::signal::unix::{signal, SignalKind};
+        let mut term = signal(SignalKind::terminate())?;
+        let mut int = signal(SignalKind::interrupt())?;
+        Ok(async move {
+            tokio::select! {
+                _ = term.recv() => {}
+                _ = int.recv() => {}
+            }
+        })
+    }
+    #[cfg(not(unix))]
+    {
+        Ok(async {
+            let _ = tokio::signal::ctrl_c().await;
+        })
+    }
+}
+
+/// Ask a worker to exit (SIGTERM lets it flush cookies), then kill it if it is
+/// still running after a short grace period.
+async fn stop_worker(child: &mut tokio::process::Child) {
+    #[cfg(unix)]
+    if let Some(pid) = child.id() {
+        // SAFETY: plain signal send to a child we still own (not yet reaped).
+        unsafe { libc::kill(pid as libc::pid_t, libc::SIGTERM) };
+        if tokio::time::timeout(Duration::from_secs(3), child.wait())
+            .await
+            .is_ok()
+        {
+            return;
+        }
+    }
+    let _ = child.kill().await;
 }
 
 async fn settle_page(page: &mut Page, wait_secs: u64, fixed: bool) {
diff --git a/crates/obscura-cli/tests/serve_shutdown.rs b/crates/obscura-cli/tests/serve_shutdown.rs
new file mode 100644
index 0000000..60a1a00
--- /dev/null
+++ b/crates/obscura-cli/tests/serve_shutdown.rs
@@ -0,0 +1,79 @@
+#![cfg(target_os = "linux")]
+
+use std::io::{Read, Write};
+use std::net::{TcpListener, TcpStream};
+use std::os::unix::process::CommandExt;
+use std::path::Path;
+use std::process::{Child, Command, Stdio};
+use std::time::{Duration, Instant};
+
+struct Server(Child);
+
+impl Drop for Server {
+    fn drop(&mut self) {
+        let _ = Command::new("kill").args(["-KILL", "--", &format!("-{}", self.0.id())])
+            .stdout(Stdio::null()).stderr(Stdio::null()).status();
+        let _ = self.0.wait();
+    }
+}
+
+fn workers(parent: u32) -> Vec<u32> {
+    let root = format!("/proc/{parent}/task");
+    let mut children = Vec::new();
+    for task in std::fs::read_dir(root).unwrap() {
+        if let Ok(ids) = std::fs::read_to_string(task.unwrap().path().join("children")) {
+            children.extend(ids.split_whitespace().map(|id| id.parse::<u32>().unwrap()));
+        }
+    }
+    children.sort_unstable();
+    children.dedup();
+    children
+}
+
+fn discovery_ok(port: u16) -> bool {
+    let Ok(mut stream) = TcpStream::connect(("127.0.0.1", port)) else { return false };
+    stream.set_read_timeout(Some(Duration::from_secs(2))).unwrap();
+    write!(stream, "GET /json/version HTTP/1.1\r\nHost: 127.0.0.1:{port}\r\nConnection: close\r\n\r\n").unwrap();
+    let mut response = String::new();
+    stream.read_to_string(&mut response).is_ok() && response.starts_with("HTTP/1.1 200")
+}
+
+#[test]
+fn balancer_signal_stops_workers_and_workers_get_max_connections() {
+    for signal in ["-TERM", "-INT"] {
+        let reservation = TcpListener::bind(("127.0.0.1", 0)).unwrap();
+        let port = reservation.local_addr().unwrap().port();
+        drop(reservation);
+        let mut server = Server(Command::new(env!("CARGO_BIN_EXE_obscura"))
+            .args(["serve", "--host", "127.0.0.1", "--port", &port.to_string(),
+                "--workers", "2", "--max-connections", "7"])
+            .process_group(0).stdout(Stdio::null()).stderr(Stdio::null()).spawn().unwrap());
+        let deadline = Instant::now() + Duration::from_secs(10);
+        while !discovery_ok(port) {
+            assert!(server.0.try_wait().unwrap().is_none(), "balancer exited");
+            assert!(Instant::now() < deadline, "balancer did not become ready");
+            std::thread::sleep(Duration::from_millis(10));
+        }
+        let pids = workers(server.0.id());
+        assert_eq!(pids.len(), 2);
+        for pid in &pids {
+            let argv = std::fs::read(format!("/proc/{pid}/cmdline")).unwrap();
+            let argv: Vec<&[u8]> = argv.split(|b| *b == 0).collect();
+            assert!(argv.windows(2).any(|w| w == [&b"--max-connections"[..], b"7"]),
+                "worker {pid} missing --max-connections 7: {argv:?}");
+        }
+
+        // Signal only the balancer, not the process group.
+        assert!(Command::new("kill").args([signal, &server.0.id().to_string()]).status().unwrap().success());
+        let deadline = Instant::now() + Duration::from_secs(10);
+        while server.0.try_wait().unwrap().is_none() {
+            assert!(Instant::now() < deadline, "balancer ignored {signal}");
+            std::thread::sleep(Duration::from_millis(10));
+        }
+        let deadline = Instant::now() + Duration::from_secs(10);
+        while pids.iter().any(|pid| Path::new(&format!("/proc/{pid}")).exists()) {
+            assert!(Instant::now() < deadline, "workers survived {signal}: {pids:?}");
+            std::thread::sleep(Duration::from_millis(10));
+        }
+    }
+}