diff --git a/curp/src/client_new/mod.rs b/curp/src/client_new/mod.rs
index e6a9e6c94..9967ff9c6 100644
--- a/curp/src/client_new/mod.rs
+++ b/curp/src/client_new/mod.rs
@@ -4,6 +4,9 @@
 /// Unary rpc client
 mod unary;
 
+/// Retry layer
+mod retry;
+
 use std::{collections::HashMap, sync::Arc};
 
 use async_trait::async_trait;
diff --git a/curp/src/client_new/retry.rs b/curp/src/client_new/retry.rs
new file mode 100644
index 000000000..29ffdd407
--- /dev/null
+++ b/curp/src/client_new/retry.rs
@@ -0,0 +1,256 @@
+#![allow(unused)] // TODO: remove
+
+use std::{marker::PhantomData, ops::SubAssign, sync::Arc, time::Duration};
+
+use async_trait::async_trait;
+use curp_external_api::cmd::Command;
+use futures::Future;
+use tracing::warn;
+
+use crate::{
+    members::ServerId,
+    rpc::{
+        connect::ConnectApi, ConfChange, CurpError, FetchClusterResponse, Member, ReadState,
+        Redirect,
+    },
+};
+
+use super::{ClientApi, LeaderStateUpdate, ProposeResponse};
+
+/// Backoff config
+#[derive(Debug, Clone)]
+pub(super) enum BackoffConfig {
+    /// A fixed delay backoff
+    Fixed,
+    /// A exponential delay backoff
+    Exponential {
+        /// Control the max delay of exponential
+        max_delay: Duration,
+    },
+}
+
+/// Retry config to control the retry policy
+#[derive(Debug, Clone)]
+pub(super) struct RetryConfig {
+    /// Backoff config
+    backoff: BackoffConfig,
+    /// Initial delay
+    delay: Duration,
+    /// Retry count
+    count: usize,
+    /// Enable jitter to randomize the delay to avoid thundering herd
+    jitter: bool,
+}
+
+/// Backoff tool
+#[derive(Debug)]
+struct Backoff {
+    /// The retry config
+    config: RetryConfig,
+    /// Current delay
+    cur_delay: Duration,
+}
+
+impl RetryConfig {
+    /// Create a fixed retry config
+    fn new_fixed(delay: Duration, count: usize, jitter: bool) -> Self {
+        assert!(count > 0, "retry count should be larger than 0");
+        Self {
+            backoff: BackoffConfig::Fixed,
+            delay,
+            count,
+            jitter,
+        }
+    }
+
+    /// Create a exponential retry config
+    fn new_exponential(delay: Duration, max_delay: Duration, count: usize, jitter: bool) -> Self {
+        assert!(count > 0, "retry count should be larger than 0");
+        Self {
+            backoff: BackoffConfig::Exponential { max_delay },
+            delay,
+            count,
+            jitter,
+        }
+    }
+
+    /// Create a backoff process
+    fn init_backoff(&self) -> Backoff {
+        Backoff {
+            config: self.clone(),
+            cur_delay: self.delay,
+        }
+    }
+}
+
+impl Backoff {
+    /// Get the next delay duration, None means the end.
+    fn next_delay(&mut self) -> Option<Duration> {
+        if self.config.count == 0 {
+            return None;
+        }
+        self.config.count.sub_assign(1);
+        let mut cur = self.cur_delay;
+        if let BackoffConfig::Exponential { max_delay } = self.config.backoff {
+            self.cur_delay = self
+                .cur_delay
+                .checked_mul(2)
+                .unwrap_or(self.cur_delay)
+                .min(max_delay);
+        }
+        #[allow(clippy::float_arithmetic)] // It is always correct.
+        if self.config.jitter {
+            // jitter algorithm will randomly pick a delay between [0.5 * delay, 1.5 * delay)
+            let per: f32 = rand::random();
+            let cur_sec = cur.as_secs_f32() * (0.5 + per);
+            cur = Duration::from_secs_f32(cur_sec);
+        }
+        Some(cur)
+    }
+}
+
+/// The retry client automatically retry the requests of the inner client api
+/// which raises the [`tonic::Status`] error
+#[derive(Debug)]
+pub(super) struct Retry<Api> {
+    /// Inner client
+    inner: Api,
+    /// Retry config
+    config: RetryConfig,
+}
+
+impl<Api> Retry<Api>
+where
+    Api: ClientApi<Error = CurpError> + LeaderStateUpdate + Send + Sync + 'static,
+{
+    /// Create a retry client
+    fn new(inner: Api, config: RetryConfig) -> Self {
+        Self { inner, config }
+    }
+
+    /// Takes a function f and run retry.
+    async fn retry<'a, R, F>(&'a self, f: impl Fn(&'a Api) -> F) -> Result<R, tonic::Status>
+    where
+        F: Future<Output = Result<R, CurpError>>,
+    {
+        let mut backoff = self.config.init_backoff();
+        while let Some(delay) = backoff.next_delay() {
+            let err = match f(&self.inner).await {
+                Ok(res) => return Ok(res),
+                Err(err) => err,
+            };
+
+            match err {
+                // some errors that should not retry
+                CurpError::Duplicated(_)
+                | CurpError::ShuttingDown(_)
+                | CurpError::InvalidConfig(_)
+                | CurpError::NodeNotExists(_)
+                | CurpError::NodeAlreadyExists(_)
+                | CurpError::LearnerNotCatchUp(_) => {
+                    return Err(tonic::Status::from(err));
+                }
+
+                // some errors that could have a retry
+                CurpError::ExpiredClientId(_)
+                | CurpError::KeyConflict(_)
+                | CurpError::RpcTransport(_)
+                | CurpError::Internal(_) => {}
+
+                // update the cluster state if got WrongClusterVersion
+                CurpError::WrongClusterVersion(_) => {
+                    // the inner client should automatically update cluster state when fetch_cluster
+                    if let Err(e) = self.inner.fetch_cluster(false).await {
+                        warn!("fetch cluster failed, error {e:?}");
+                    }
+                }
+
+                // update the leader state if got Redirect
+                CurpError::Redirect(Redirect { leader_id, term }) => {
+                    self.inner.update_leader(leader_id, term);
+                }
+            }
+
+            warn!("retry on {} seconds later", delay.as_secs_f32());
+            tokio::time::sleep(delay).await;
+        }
+
+        Err(tonic::Status::deadline_exceeded("request timeout"))
+    }
+}
+
+#[async_trait]
+impl<Api> ClientApi for Retry<Api>
+where
+    Api: ClientApi<Error = CurpError> + LeaderStateUpdate + Send + Sync + 'static,
+{
+    /// The client error
+    type Error = tonic::Status;
+
+    /// Inherit the command type
+    type Cmd = Api::Cmd;
+
+    /// Get the local connection when the client is on the server node.
+    fn local_connect(&self) -> Option<Arc<dyn ConnectApi>> {
+        self.inner.local_connect()
+    }
+
+    /// Send propose to the whole cluster, `use_fast_path` set to `false` to fallback into ordered
+    /// requests (event the requests are commutative).
+    async fn propose(
+        &self,
+        cmd: &Self::Cmd,
+        use_fast_path: bool,
+    ) -> Result<ProposeResponse<Self::Cmd>, tonic::Status> {
+        self.retry::<_, _>(|client| async move { client.propose(cmd, use_fast_path).await })
+            .await
+    }
+
+    /// Send propose configuration changes to the cluster
+    async fn propose_conf_change(
+        &self,
+        changes: Vec<ConfChange>,
+    ) -> Result<Vec<Member>, tonic::Status> {
+        self.retry::<_, _>(|client| {
+            let changes_c = changes.clone();
+            async move { client.propose_conf_change(changes_c).await }
+        })
+        .await
+    }
+
+    /// Send propose to shutdown cluster
+    async fn propose_shutdown(&self) -> Result<(), tonic::Status> {
+        self.retry::<_, _>(|client| async move { client.propose_shutdown().await })
+            .await
+    }
+
+    /// Send propose to publish a node id and name
+    async fn propose_publish(
+        &self,
+        node_id: ServerId,
+        node_name: String,
+    ) -> Result<(), Self::Error> {
+        self.retry::<_, _>(|client| {
+            let name_c = node_name.clone();
+            async move { client.propose_publish(node_id, name_c).await }
+        })
+        .await
+    }
+
+    /// Send fetch read state from leader
+    async fn fetch_read_state(&self, cmd: &Self::Cmd) -> Result<ReadState, tonic::Status> {
+        self.retry::<_, _>(|client| async move { client.fetch_read_state(cmd).await })
+            .await
+    }
+
+    /// Send fetch cluster requests to all servers (That's because initially, we didn't
+    /// know who the leader is.)
+    /// Note: The fetched cluster may still be outdated if `linearizable` is false
+    async fn fetch_cluster(
+        &self,
+        linearizable: bool,
+    ) -> Result<FetchClusterResponse, tonic::Status> {
+        self.retry::<_, _>(|client| async move { client.fetch_cluster(linearizable).await })
+            .await
+    }
+}