diff --git a/vhost/CHANGELOG.md b/vhost/CHANGELOG.md
index e17cee7c..9dde328c 100644
--- a/vhost/CHANGELOG.md
+++ b/vhost/CHANGELOG.md
@@ -3,6 +3,8 @@
 ## [Unreleased]
 
 ### Added
+- [[#369]](https://github.com/rust-vmm/vhost/pull/369) Add `GpuFrontend` API and `VhostUserFrontend::set_gpu_socket()` for frontend-side GPU socket communication.
+
 ### Changed
 ### Deprecated
 ### Fixed
diff --git a/vhost/src/vhost_user/connection.rs b/vhost/src/vhost_user/connection.rs
index d9859182..1ffeb165 100644
--- a/vhost/src/vhost_user/connection.rs
+++ b/vhost/src/vhost_user/connection.rs
@@ -424,6 +424,9 @@ impl<H: MsgHeader> Endpoint<H> {
         &mut self,
         buf_size: usize,
     ) -> Result<(usize, Vec<u8>, Option<Vec<File>>)> {
+        if buf_size > H::MAX_MSG_SIZE {
+            return Err(Error::OversizedMsg);
+        }
         let mut buf = vec![0u8; buf_size];
         let (bytes, files) = {
             let mut iovs = [iovec {
diff --git a/vhost/src/vhost_user/frontend.rs b/vhost/src/vhost_user/frontend.rs
index 851beff4..6f1cf3ef 100644
--- a/vhost/src/vhost_user/frontend.rs
+++ b/vhost/src/vhost_user/frontend.rs
@@ -59,6 +59,9 @@ pub trait VhostUserFrontend: VhostBackend {
     /// Setup backend communication channel.
     fn set_backend_request_fd(&mut self, fd: &dyn AsRawFd) -> Result<()>;
 
+    /// Setup GPU protocol socket.
+    fn set_gpu_socket(&mut self, fd: &dyn AsRawFd) -> Result<()>;
+
     /// Retrieve a given dma-buf fd from a given backend
     fn get_shared_object(&mut self, uuid: &VhostUserSharedMsg) -> Result<File>;
 
@@ -506,6 +509,13 @@ impl VhostUserFrontend for Frontend {
         node.wait_for_ack(&hdr).map_err(|e| e.into())
     }
 
+    fn set_gpu_socket(&mut self, fd: &dyn AsRawFd) -> Result<()> {
+        let mut node = self.node();
+        let fds = [fd.as_raw_fd()];
+        let hdr = node.send_request_header(FrontendReq::GPU_SET_SOCKET, Some(&fds))?;
+        node.wait_for_ack(&hdr).map_err(|e| e.into())
+    }
+
     fn get_shared_object(&mut self, uuid: &VhostUserSharedMsg) -> Result<File> {
         let mut node = self.node();
         node.check_proto_feature(VhostUserProtocolFeatures::SHARED_OBJECT)?;
diff --git a/vhost/src/vhost_user/gpu_frontend_req.rs b/vhost/src/vhost_user/gpu_frontend_req.rs
new file mode 100644
index 00000000..2889340c
--- /dev/null
+++ b/vhost/src/vhost_user/gpu_frontend_req.rs
@@ -0,0 +1,226 @@
+// Copyright (C) 2026 Red Hat, Inc.
+// SPDX-License-Identifier: Apache-2.0
+
+use std::fs::File;
+use std::os::fd::{AsRawFd, RawFd};
+use std::os::unix::net::UnixStream;
+use std::sync::{Arc, Mutex, MutexGuard};
+use std::{io, mem};
+
+use vm_memory::ByteValued;
+
+use crate::vhost_user::connection::Endpoint;
+use crate::vhost_user::gpu_message::{
+    GpuBackendReq, VhostUserGpuHeaderFlag, VhostUserGpuMsgHeader,
+};
+
+struct FrontendInternal {
+    sock: Endpoint<VhostUserGpuMsgHeader<GpuBackendReq>>,
+    error: Option<i32>,
+}
+
+impl FrontendInternal {
+    fn check_state(&self) -> io::Result<()> {
+        match self.error {
+            Some(e) => Err(io::Error::from_raw_os_error(e)),
+            None => Ok(()),
+        }
+    }
+
+    fn send_response<T: ByteValued>(
+        &mut self,
+        request: GpuBackendReq,
+        body: &T,
+        fds: Option<&[RawFd]>,
+    ) -> io::Result<()> {
+        self.check_state()?;
+
+        let len = mem::size_of::<T>();
+        let hdr =
+            VhostUserGpuMsgHeader::new(request, VhostUserGpuHeaderFlag::REPLY.bits(), len as u32);
+        self.sock
+            .send_message(&hdr, body, fds)
+            .map_err(io::Error::other)?;
+        Ok(())
+    }
+}
+
+/// Proxy for receiving messages from the backend and sending responses
+/// over the socket obtained from VHOST_USER_GPU_SET_SOCKET.
+/// The protocol is documented here: <https://www.qemu.org/docs/master/interop/vhost-user-gpu.html>
+#[derive(Clone)]
+pub struct GpuFrontend {
+    // underlying Unix domain socket for communication
+    node: Arc<Mutex<FrontendInternal>>,
+}
+
+impl GpuFrontend {
+    fn new(ep: Endpoint<VhostUserGpuMsgHeader<GpuBackendReq>>) -> Self {
+        Self {
+            node: Arc::new(Mutex::new(FrontendInternal {
+                sock: ep,
+                error: None,
+            })),
+        }
+    }
+
+    fn node(&self) -> MutexGuard<'_, FrontendInternal> {
+        self.node.lock().unwrap()
+    }
+
+    /// Create a new instance from a `UnixStream` object.
+    pub fn from_stream(sock: UnixStream) -> Self {
+        Self::new(Endpoint::<VhostUserGpuMsgHeader<GpuBackendReq>>::from_stream(sock))
+    }
+
+    /// Read a GPU protocol message from the backend.
+    /// Returns (message_type, payload_bytes, optional_fds).
+    /// File descriptors are present for DMABUF_SCANOUT and DMABUF_SCANOUT2 messages.
+    pub fn read_message(&self) -> io::Result<(GpuBackendReq, Vec<u8>, Option<Vec<File>>)> {
+        let mut node = self.node();
+        node.check_state()?;
+
+        let (hdr, files) = node.sock.recv_header().map_err(io::Error::other)?;
+        let request = hdr.get_code().map_err(io::Error::other)?;
+        let size = hdr.get_size() as usize;
+
+        let payload = if size > 0 {
+            let (bytes, mut payload, _) =
+                node.sock.recv_into_buf(size).map_err(io::Error::other)?;
+            payload.truncate(bytes);
+            payload
+        } else {
+            Vec::new()
+        };
+
+        Ok((request, payload, files))
+    }
+
+    /// Set error state for the endpoint.
+    pub fn set_failed(&self, error: i32) {
+        self.node().error = Some(error);
+    }
+
+    /// Send a typed response to a GPU protocol request.
+    pub fn send_response<T: ByteValued>(
+        &self,
+        request: GpuBackendReq,
+        response: &T,
+    ) -> io::Result<()> {
+        let mut node = self.node();
+        node.send_response(request, response, None)?;
+        Ok(())
+    }
+}
+
+impl AsRawFd for GpuFrontend {
+    fn as_raw_fd(&self) -> RawFd {
+        self.node.lock().unwrap().sock.as_raw_fd()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::vhost_user::gpu_message::{
+        VhostUserGpuEdidRequest, VirtioGpuDisplayOne, VirtioGpuRect, VirtioGpuRespDisplayInfo,
+    };
+    use std::mem::size_of;
+    use std::thread;
+
+    fn frontend_backend_pair() -> (GpuFrontend, Endpoint<VhostUserGpuMsgHeader<GpuBackendReq>>) {
+        let (fe, be) = UnixStream::pair().unwrap();
+        (GpuFrontend::from_stream(fe), Endpoint::from_stream(be))
+    }
+
+    #[test]
+    fn test_read_message_with_payload() {
+        let (frontend, mut backend) = frontend_backend_pair();
+
+        let request = VhostUserGpuEdidRequest { scanout_id: 1 };
+        let sender = thread::spawn(move || {
+            let hdr = VhostUserGpuMsgHeader::new(
+                GpuBackendReq::GET_EDID,
+                0,
+                size_of::<VhostUserGpuEdidRequest>() as u32,
+            );
+            backend.send_message(&hdr, &request, None).unwrap();
+        });
+
+        let (req, payload, fds) = frontend.read_message().unwrap();
+        assert_eq!(req, GpuBackendReq::GET_EDID);
+        assert_eq!(payload.len(), size_of::<VhostUserGpuEdidRequest>());
+        assert!(fds.is_none());
+
+        sender.join().unwrap();
+    }
+
+    #[test]
+    fn test_read_message_no_payload() {
+        let (frontend, mut backend) = frontend_backend_pair();
+
+        let sender = thread::spawn(move || {
+            let hdr = VhostUserGpuMsgHeader::new(GpuBackendReq::GET_DISPLAY_INFO, 0, 0);
+            backend.send_header(&hdr, None).unwrap();
+        });
+
+        let (req, payload, fds) = frontend.read_message().unwrap();
+        assert_eq!(req, GpuBackendReq::GET_DISPLAY_INFO);
+        assert!(payload.is_empty());
+        assert!(fds.is_none());
+
+        sender.join().unwrap();
+    }
+
+    #[test]
+    fn test_roundtrip_display_info() {
+        let (frontend, mut backend) = frontend_backend_pair();
+
+        let mut expected = VirtioGpuRespDisplayInfo::default();
+        expected.pmodes[0] = VirtioGpuDisplayOne {
+            r: VirtioGpuRect {
+                x: 0,
+                y: 0,
+                width: 1920,
+                height: 1080,
+            },
+            enabled: 1,
+            flags: 0,
+        };
+
+        let sender = thread::spawn(move || {
+            let hdr = VhostUserGpuMsgHeader::new(GpuBackendReq::GET_DISPLAY_INFO, 0, 0);
+            backend.send_header(&hdr, None).unwrap();
+
+            let (reply_hdr, body, _fds) = backend.recv_body::<VirtioGpuRespDisplayInfo>().unwrap();
+            assert!(reply_hdr.is_reply());
+            assert_eq!(
+                reply_hdr.get_code().unwrap(),
+                GpuBackendReq::GET_DISPLAY_INFO
+            );
+            body
+        });
+
+        let (req, _payload, _fds) = frontend.read_message().unwrap();
+        frontend.send_response(req, &expected).unwrap();
+
+        let body = sender.join().unwrap();
+        assert_eq!(body, expected);
+    }
+
+    #[test]
+    fn test_set_failed() {
+        let (fe, _be) = UnixStream::pair().unwrap();
+        let frontend = GpuFrontend::from_stream(fe);
+
+        frontend.set_failed(libc::EAGAIN);
+
+        assert!(frontend.read_message().is_err());
+        assert!(frontend
+            .send_response(
+                GpuBackendReq::GET_DISPLAY_INFO,
+                &VirtioGpuRespDisplayInfo::default()
+            )
+            .is_err());
+    }
+}
diff --git a/vhost/src/vhost_user/mod.rs b/vhost/src/vhost_user/mod.rs
index 7012222f..28eddfa3 100644
--- a/vhost/src/vhost_user/mod.rs
+++ b/vhost/src/vhost_user/mod.rs
@@ -52,8 +52,10 @@ mod backend_req;
 #[cfg(feature = "vhost-user-backend")]
 pub use self::backend_req::Backend;
 mod gpu_backend_req;
+mod gpu_frontend_req;
 pub mod gpu_message;
 pub use self::gpu_backend_req::GpuBackend;
+pub use self::gpu_frontend_req::GpuFrontend;
 
 /// Errors for vhost-user operations
 #[derive(Debug)]