diff --git a/src/stream/tcb.rs b/src/stream/tcb.rs
index abff5c1..7b2bc7a 100644
--- a/src/stream/tcb.rs
+++ b/src/stream/tcb.rs
@@ -63,6 +63,8 @@ pub(crate) struct Tcb {
     seq: SeqNum,
     ack: SeqNum,
     mtu: u16,
+    /// The largest segment the peer announced in the MSS option of its SYN, unbounded without one.
+    peer_mss: u16,
     last_received_ack: SeqNum,
     send_window: u32,
     send_window_shift: u8,
@@ -125,6 +127,7 @@ impl Tcb {
             seq: seq.into(),
             ack,
             mtu,
+            peer_mss: u16::MAX,
             last_received_ack: seq.into(),
             send_window: peer_window as u32,
             send_window_shift,
@@ -144,10 +147,19 @@ impl Tcb {
         }
     }
 
+    /// RFC 9293 §3.7.1: a sender never sends a segment larger than the peer announced. Below 88,
+    /// the floor Linux applies (TCP_MIN_MSS), the option is not followed, so a peer cannot ask for
+    /// its data a few bytes at a time.
+    pub(super) fn set_peer_mss(&mut self, mss: u16) {
+        self.peer_mss = mss.max(88);
+    }
+
     pub fn calculate_payload_max_len(&self, ip_header_size: usize, tcp_header_size: usize) -> usize {
         let send_window = self.get_send_window() as usize;
         let mtu = self.get_mtu() as usize;
-        std::cmp::min(send_window, mtu.saturating_sub(ip_header_size + tcp_header_size))
+        // The MSS counts the payload and the TCP options of a segment, but not the 20-byte base header.
+        let mss = (self.peer_mss as usize).saturating_sub(tcp_header_size.saturating_sub(20));
+        send_window.min(mtu.saturating_sub(ip_header_size + tcp_header_size)).min(mss)
     }
 
     pub fn update_duplicate_ack_count(&mut self, rcvd_ack: SeqNum) {
diff --git a/src/stream/tcp.rs b/src/stream/tcp.rs
index 212b7a0..ab6c62b 100644
--- a/src/stream/tcp.rs
+++ b/src/stream/tcp.rs
@@ -196,7 +196,7 @@ impl IpStackTcpStream {
                 })
             })
             .flatten();
-        let tcb = Tcb::new(
+        let mut tcb = Tcb::new(
             SeqNum(tcp.sequence_number),
             tcp.window_size,
             peer_window_shift,
@@ -207,6 +207,15 @@ impl IpStackTcpStream {
             config.rto,
             config.max_retransmit_count,
         );
+        // Like the window scale, the MSS option is read only from a segment carrying SYN.
+        if tcp.syn
+            && let Some(mss) = tcp.options_iterator().flatten().find_map(|option| match option {
+                TcpOptionElement::MaximumSegmentSize(mss) => Some(mss),
+                _ => None,
+            })
+        {
+            tcb.set_peer_mss(mss);
+        }
         let tuple = NetworkTuple::new(src_addr, dst_addr, true);
         if !tcp.syn {
             if !tcp.rst
@@ -1575,6 +1584,27 @@ mod tests {
         assert_eq!(tcb.get_send_window(), 40_000 << 14);
     }
 
+    /// RFC 9293 §3.7.1: a segment never carries more payload than the MSS the peer announced in its
+    /// SYN, and an announced MSS below 88 (Linux's TCP_MIN_MSS) is read as 88.
+    #[tokio::test]
+    async fn segments_stay_within_the_peers_mss() {
+        for (announced, expected) in [(600, 600), (100, 100), (40, 88)] {
+            let (mut stream, syn_ack, mut up_rx) = open(u16::MAX, &[TcpOptionElement::MaximumSegmentSize(announced)]).await;
+            let peer_seq = SeqNum(syn_ack.acknowledgment_number);
+            feed(&stream, &syn_ack, peer_seq, u16::MAX, &[1; 4]);
+            up_rx.recv().await.unwrap();
+
+            let (_, waker) = recording_waker();
+            let mut cx = Context::from_waker(&waker);
+            let written = match Pin::new(&mut stream).poll_write(&mut cx, &[7; 1460]) {
+                Poll::Ready(Ok(n)) => n,
+                other => panic!("the write did not complete: {other:?}"),
+            };
+            assert_eq!(written, expected, "peer MSS {announced}: wrong write size");
+            assert_eq!(up_rx.recv().await.unwrap().payload.map(|p| p.len()), Some(expected));
+        }
+    }
+
     /// A closed peer window holds the writer, and the segment that reopens it wakes the writer, both
     /// when it completes the handshake and when it arrives on an established connection.
     #[tokio::test]