diff --git a/core/archipelago/src/fips/anchors.rs b/core/archipelago/src/fips/anchors.rs
index c92a5b48..9362d15b 100644
--- a/core/archipelago/src/fips/anchors.rs
+++ b/core/archipelago/src/fips/anchors.rs
@@ -216,6 +216,44 @@ pub struct ApplyResult {
     pub message: String,
 }
 
+/// FIPS UDP transport port (matches `transports.udp.bind_addr` in the generated
+/// `fips.yaml`). Direct peer links dial this, NOT the HTTP/LAN messaging port.
+const FIPS_UDP_PORT: u16 = 8668;
+
+/// Build transient seed-anchor entries that dial LAN-discovered federation peers
+/// directly over their FIPS UDP transport. For each peer the registry knows both
+/// a LAN socket address AND a FIPS npub for, point a `udp` anchor at
+/// `<lan-ip>:8668`. This lets co-located federation nodes form a DIRECT FIPS link
+/// instead of depending on the global anchor's spanning tree to route between
+/// them (the cause of every dial falling back to Tor when the anchor link flaps).
+///
+/// This is FIPS's own UDP transport over the LAN — not Tailscale, not the LAN
+/// HTTP messaging port. NOT persisted to `seed-anchors.json`: recomputed each
+/// apply tick from live LAN discovery, so a peer's changing IP self-corrects and
+/// stale entries never accumulate. `fipsctl connect` is idempotent, so
+/// re-applying just keeps the link warm.
+pub fn lan_fips_anchors(peers: &[crate::transport::PeerRecord]) -> Vec<SeedAnchor> {
+    let mut out = Vec::new();
+    for p in peers {
+        let (Some(lan), Some(npub)) = (p.lan_address.as_deref(), p.fips_npub.as_deref()) else {
+            continue;
+        };
+        // lan_address is the peer's HTTP/LAN socket ("ip:port"); reuse only its IP
+        // and target the FIPS UDP port. SocketAddr::new(...).to_string() formats
+        // IPv6 with brackets correctly.
+        let Ok(sa) = lan.parse::<std::net::SocketAddr>() else {
+            continue;
+        };
+        out.push(SeedAnchor {
+            npub: npub.to_string(),
+            address: std::net::SocketAddr::new(sa.ip(), FIPS_UDP_PORT).to_string(),
+            transport: "udp".to_string(),
+            label: "LAN federation peer (direct FIPS)".to_string(),
+        });
+    }
+    out
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
diff --git a/core/archipelago/src/server.rs b/core/archipelago/src/server.rs
index ece04176..32c60f1e 100644
--- a/core/archipelago/src/server.rs
+++ b/core/archipelago/src/server.rs
@@ -355,6 +355,9 @@ impl Server {
         }
 
         // Initialize transport router (unified routing: mesh > lan > tor)
+        // Hoisted so the FIPS seed-anchor loop below can auto-peer LAN-discovered
+        // federation peers directly over FIPS (see that loop).
+        let mut fips_peer_registry: Option<std::sync::Arc<crate::transport::PeerRegistry>> = None;
         {
             let data_dir = config.data_dir.clone();
             let did =
@@ -368,6 +371,7 @@ impl Server {
             match crate::transport::PeerRegistry::load(&data_dir).await {
                 Ok(registry) => {
                     let registry = std::sync::Arc::new(registry);
+                    fips_peer_registry = Some(registry.clone());
                     let mut transports: Vec<Box<dyn crate::transport::NodeTransport>> = Vec::new();
 
                     // Tor transport (always register — availability checked dynamically)
@@ -641,6 +645,7 @@ impl Server {
         // onboarding before we start dialing.
         {
             let data_dir = config.data_dir.clone();
+            let fips_peer_registry = fips_peer_registry.clone();
             tokio::spawn(async move {
                 tokio::time::sleep(Duration::from_secs(30)).await;
                 let mut interval = tokio::time::interval(Duration::from_secs(300));
@@ -655,6 +660,23 @@ impl Server {
                             tracing::debug!("Seed-anchor apply: load failed (non-fatal): {}", e)
                         }
                     }
+
+                    // Auto-peer federation nodes we've discovered on the LAN
+                    // directly over FIPS, so co-located peers don't depend on the
+                    // (often flaky) global anchor's spanning tree to route to each
+                    // other. For every peer the registry knows both a LAN address
+                    // AND a FIPS npub for, dial it on its FIPS UDP transport port
+                    // (8668) at its LAN IP. This is FIPS's own transport over the
+                    // LAN — NOT Tailscale, NOT the HTTP/LAN messaging port. Pure
+                    // FIPS. `fipsctl connect` is idempotent, so re-applying every
+                    // tick just keeps the direct link warm; unknown/remote peers
+                    // (no LAN address) are left to the anchor as before.
+                    if let Some(reg) = fips_peer_registry.as_ref() {
+                        let direct = crate::fips::anchors::lan_fips_anchors(&reg.all_peers().await);
+                        if !direct.is_empty() {
+                            let _ = crate::fips::anchors::apply(&direct).await;
+                        }
+                    }
                 }
             });
         }