diff --git a/crates/switchyard-server/src/lib.rs b/crates/switchyard-server/src/lib.rs
index eea1c7a2e..91a367011 100644
--- a/crates/switchyard-server/src/lib.rs
+++ b/crates/switchyard-server/src/lib.rs
@@ -479,12 +479,18 @@ async fn stamp_request_start(mut request: HttpRequest, next: Next) -> Response {
     next.run(request).await
 }
 
-/// Builds an Axum router for the supported LLM wire formats.
+/// Builds an Axum router containing only the primary LLM inference endpoints.
+///
+/// The registered paths are `/v1/chat/completions`, `/v1/messages`, and `/v1/responses`.
+/// The router excludes operational, discovery, auxiliary, and fallback proxy routes so an
+/// embedding application can mount those capabilities under its own policies.
+pub fn build_llm_router(state: ServerState) -> Router {
+    finish_router(primary_llm_routes(), state)
+}
+
+/// Builds the full Axum router used by the standalone Switchyard server.
 pub fn build_switchyard_router(state: ServerState) -> Router {
-    let mut router = Router::new()
-        .route("/v1/chat/completions", post(openai_chat_completions))
-        .route("/v1/messages", post(anthropic_messages))
-        .route("/v1/responses", post(openai_responses))
+    let mut router = primary_llm_routes()
         .route("/v1/decision", post(decision))
         .route("/v1/messages/count_tokens", post(anthropic_count_tokens))
         .route(
@@ -500,8 +506,20 @@ pub fn build_switchyard_router(state: ServerState) -> Router {
     if state.routing_log.is_some() {
         router = router.route("/v1/routing/session-stats", get(get_session_stats));
     }
+    finish_router(router.fallback(proxy_unmatched), state)
+}
+
+// Keeps the embedded and standalone servers on the same primary route definitions.
+fn primary_llm_routes() -> Router<ServerState> {
+    Router::new()
+        .route("/v1/chat/completions", post(openai_chat_completions))
+        .route("/v1/messages", post(anthropic_messages))
+        .route("/v1/responses", post(openai_responses))
+}
+
+// Applies the serving limits and request timing shared by both public router constructors.
+fn finish_router(router: Router<ServerState>, state: ServerState) -> Router {
     router
-        .fallback(proxy_unmatched)
         .layer(DefaultBodyLimit::max(DEFAULT_MAX_REQUEST_BODY_BYTES))
         // `layer` only wraps routes registered before it, so this stays last.
         .layer(axum::middleware::from_fn(stamp_request_start))
diff --git a/crates/switchyard-server/tests/server.rs b/crates/switchyard-server/tests/server.rs
index 4fca26f02..f385f7e3c 100644
--- a/crates/switchyard-server/tests/server.rs
+++ b/crates/switchyard-server/tests/server.rs
@@ -25,7 +25,9 @@ use switchyard_llm_client::{
 use switchyard_protocol::ModelId;
 use switchyard_protocol::RoutedLlmClient;
 use switchyard_server::config::load_server_state;
-use switchyard_server::{DEFAULT_MAX_REQUEST_BODY_BYTES, ServerState, build_switchyard_router};
+use switchyard_server::{
+    DEFAULT_MAX_REQUEST_BODY_BYTES, ServerState, build_llm_router, build_switchyard_router,
+};
 use tokio::net::TcpListener;
 use tokio::sync::Mutex;
 use tokio::task::JoinHandle;
@@ -528,6 +530,42 @@ async fn test_app(routes: &[(&str, &[&str])]) -> TestResult<(MockUpstream, Route
     Ok((upstream, app))
 }
 
+// Embedders expose only the three primary inference endpoints and own every other route.
+#[tokio::test]
+async fn llm_router_exposes_only_primary_llm_endpoints() -> TestResult {
+    let state = random_state("http://127.0.0.1:1/v1", &[(ROUTE_MODEL, &["model/weak"])])?;
+    let app = build_llm_router(state);
+
+    for path in ["/v1/chat/completions", "/v1/messages", "/v1/responses"] {
+        assert_eq!(
+            send(&app, "GET", path, None).await?.status,
+            StatusCode::METHOD_NOT_ALLOWED,
+            "{path} should be registered"
+        );
+    }
+
+    for path in [
+        "/v1/decision",
+        "/v1/messages/count_tokens",
+        "/v1/responses/input_tokens",
+        "/v1/responses/compact",
+        "/v1/models",
+        "/v1/stats",
+        "/v1/stats/reset",
+        "/v1/routing/session-stats",
+        "/metrics",
+        "/health",
+        "/future/provider/endpoint",
+    ] {
+        assert_eq!(
+            send(&app, "POST", path, None).await?.status,
+            StatusCode::NOT_FOUND,
+            "{path} should not be registered"
+        );
+    }
+    Ok(())
+}
+
 fn empty_token_totals() -> Value {
     json!({
         "prompt": 0,