diff --git a/Cargo.lock b/Cargo.lock
index 83138eac..d5d41b03 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -840,7 +840,6 @@ dependencies = [
  "cuprate-test-utils",
  "cuprate-types",
  "cuprate-wire",
- "dashmap",
  "futures",
  "indexmap",
  "monero-serai",
diff --git a/binaries/cuprated/src/blockchain/syncer.rs b/binaries/cuprated/src/blockchain/syncer.rs
index 913c9834..69ad3303 100644
--- a/binaries/cuprated/src/blockchain/syncer.rs
+++ b/binaries/cuprated/src/blockchain/syncer.rs
@@ -12,7 +12,7 @@ use tracing::instrument;
 use cuprate_consensus::{BlockChainContext, BlockChainContextRequest, BlockChainContextResponse};
 use cuprate_p2p::{
     block_downloader::{BlockBatch, BlockDownloaderConfig, ChainSvcRequest, ChainSvcResponse},
-    NetworkInterface,
+    NetworkInterface, PeerSetRequest, PeerSetResponse,
 };
 use cuprate_p2p_core::ClearNet;
 
@@ -28,15 +28,11 @@ pub enum SyncerError {
 }
 
 /// The syncer tasks that makes sure we are fully synchronised with our connected peers.
-#[expect(
-    clippy::significant_drop_tightening,
-    reason = "Client pool which will be removed"
-)]
 #[instrument(level = "debug", skip_all)]
 pub async fn syncer<C, CN>(
     mut context_svc: C,
     our_chain: CN,
-    clearnet_interface: NetworkInterface<ClearNet>,
+    mut clearnet_interface: NetworkInterface<ClearNet>,
     incoming_block_batch_tx: mpsc::Sender<BlockBatch>,
     stop_current_block_downloader: Arc<Notify>,
     block_downloader_config: BlockDownloaderConfig,
@@ -67,8 +63,6 @@ where
         unreachable!();
     };
 
-    let client_pool = clearnet_interface.client_pool();
-
     tracing::debug!("Waiting for new sync info in top sync channel");
 
     loop {
@@ -79,9 +73,20 @@ where
         check_update_blockchain_context(&mut context_svc, &mut blockchain_ctx).await?;
         let raw_blockchain_context = blockchain_ctx.unchecked_blockchain_context();
 
-        if !client_pool.contains_client_with_more_cumulative_difficulty(
-            raw_blockchain_context.cumulative_difficulty,
-        ) {
+        let PeerSetResponse::MostPoWSeen {
+            cumulative_difficulty,
+            ..
+        } = clearnet_interface
+            .peer_set()
+            .ready()
+            .await?
+            .call(PeerSetRequest::MostPoWSeen)
+            .await?
+        else {
+            unreachable!();
+        };
+
+        if cumulative_difficulty <= raw_blockchain_context.cumulative_difficulty {
             continue;
         }
 
diff --git a/binaries/cuprated/src/txpool/dandelion.rs b/binaries/cuprated/src/txpool/dandelion.rs
index d791b62d..00d9f5a5 100644
--- a/binaries/cuprated/src/txpool/dandelion.rs
+++ b/binaries/cuprated/src/txpool/dandelion.rs
@@ -59,7 +59,7 @@ pub fn dandelion_router(clear_net: NetworkInterface<ClearNet>) -> ConcreteDandel
         diffuse_service::DiffuseService {
             clear_net_broadcast_service: clear_net.broadcast_svc(),
         },
-        stem_service::OutboundPeerStream { clear_net },
+        stem_service::OutboundPeerStream::new(clear_net),
         DANDELION_CONFIG,
     )
 }
diff --git a/binaries/cuprated/src/txpool/dandelion/stem_service.rs b/binaries/cuprated/src/txpool/dandelion/stem_service.rs
index 5c0ba65e..2debfd42 100644
--- a/binaries/cuprated/src/txpool/dandelion/stem_service.rs
+++ b/binaries/cuprated/src/txpool/dandelion/stem_service.rs
@@ -1,14 +1,15 @@
 use std::{
+    future::Future,
     pin::Pin,
-    task::{Context, Poll},
+    task::{ready, Context, Poll},
 };
 
 use bytes::Bytes;
-use futures::Stream;
+use futures::{future::BoxFuture, FutureExt, Stream};
 use tower::Service;
 
 use cuprate_dandelion_tower::{traits::StemRequest, OutboundPeer};
-use cuprate_p2p::{ClientPoolDropGuard, NetworkInterface};
+use cuprate_p2p::{ClientDropGuard, NetworkInterface, PeerSetRequest, PeerSetResponse};
 use cuprate_p2p_core::{
     client::{Client, InternalPeerID},
     ClearNet, NetworkZone, PeerRequest, ProtocolRequest,
@@ -19,7 +20,17 @@ use crate::{p2p::CrossNetworkInternalPeerId, txpool::dandelion::DandelionTx};
 
 /// The dandelion outbound peer stream.
 pub struct OutboundPeerStream {
-    pub clear_net: NetworkInterface<ClearNet>,
+    clear_net: NetworkInterface<ClearNet>,
+    state: OutboundPeerStreamState,
+}
+
+impl OutboundPeerStream {
+    pub const fn new(clear_net: NetworkInterface<ClearNet>) -> Self {
+        Self {
+            clear_net,
+            state: OutboundPeerStreamState::Standby,
+        }
+    }
 }
 
 impl Stream for OutboundPeerStream {
@@ -28,23 +39,49 @@ impl Stream for OutboundPeerStream {
         tower::BoxError,
     >;
 
-    fn poll_next(self: Pin<&mut Self>, _: &mut Context<'_>) -> Poll<Option<Self::Item>> {
-        // TODO: make the outbound peer choice random.
-        Poll::Ready(Some(Ok(self
-            .clear_net
-            .client_pool()
-            .outbound_client()
-            .map_or(OutboundPeer::Exhausted, |client| {
-                OutboundPeer::Peer(
-                    CrossNetworkInternalPeerId::ClearNet(client.info.id),
-                    StemPeerService(client),
-                )
-            }))))
+    fn poll_next(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
+        loop {
+            match &mut self.state {
+                OutboundPeerStreamState::Standby => {
+                    let peer_set = self.clear_net.peer_set();
+                    let res = ready!(peer_set.poll_ready(cx));
+
+                    self.state = OutboundPeerStreamState::AwaitingPeer(
+                        peer_set.call(PeerSetRequest::StemPeer).boxed(),
+                    );
+                }
+                OutboundPeerStreamState::AwaitingPeer(fut) => {
+                    let res = ready!(fut.poll_unpin(cx));
+
+                    return Poll::Ready(Some(res.map(|res| {
+                        let PeerSetResponse::StemPeer(stem_peer) = res else {
+                            unreachable!()
+                        };
+
+                        match stem_peer {
+                            Some(peer) => OutboundPeer::Peer(
+                                CrossNetworkInternalPeerId::ClearNet(peer.info.id),
+                                StemPeerService(peer),
+                            ),
+                            None => OutboundPeer::Exhausted,
+                        }
+                    })));
+                }
+            }
+        }
     }
 }
 
+/// The state of the [`OutboundPeerStream`].
+enum OutboundPeerStreamState {
+    /// Standby state.
+    Standby,
+    /// Awaiting a response from the peer-set.
+    AwaitingPeer(BoxFuture<'static, Result<PeerSetResponse<ClearNet>, tower::BoxError>>),
+}
+
 /// The stem service, used to send stem txs.
-pub struct StemPeerService<N: NetworkZone>(ClientPoolDropGuard<N>);
+pub struct StemPeerService<N: NetworkZone>(ClientDropGuard<N>);
 
 impl<N: NetworkZone> Service<StemRequest<DandelionTx>> for StemPeerService<N> {
     type Response = <Client<N> as Service<PeerRequest>>::Response;
diff --git a/p2p/p2p-core/src/client.rs b/p2p/p2p-core/src/client.rs
index 73b33ba6..f2fde670 100644
--- a/p2p/p2p-core/src/client.rs
+++ b/p2p/p2p-core/src/client.rs
@@ -27,9 +27,11 @@ mod connector;
 pub mod handshaker;
 mod request_handler;
 mod timeout_monitor;
+mod weak;
 
 pub use connector::{ConnectRequest, Connector};
 pub use handshaker::{DoHandshakeRequest, HandshakeError, HandshakerBuilder};
+pub use weak::WeakClient;
 
 /// An internal identifier for a given peer, will be their address if known
 /// or a random u128 if not.
@@ -128,6 +130,17 @@ impl<Z: NetworkZone> Client<Z> {
         }
         .into()
     }
+
+    /// Create a [`WeakClient`] for this [`Client`].
+    pub fn downgrade(&self) -> WeakClient<Z> {
+        WeakClient {
+            info: self.info.clone(),
+            connection_tx: self.connection_tx.downgrade(),
+            semaphore: self.semaphore.clone(),
+            permit: None,
+            error: self.error.clone(),
+        }
+    }
 }
 
 impl<Z: NetworkZone> Service<PeerRequest> for Client<Z> {
diff --git a/p2p/p2p-core/src/client/weak.rs b/p2p/p2p-core/src/client/weak.rs
new file mode 100644
index 00000000..90f25dd6
--- /dev/null
+++ b/p2p/p2p-core/src/client/weak.rs
@@ -0,0 +1,114 @@
+use std::task::{ready, Context, Poll};
+
+use futures::channel::oneshot;
+use tokio::sync::{mpsc, OwnedSemaphorePermit};
+use tokio_util::sync::PollSemaphore;
+use tower::Service;
+
+use cuprate_helper::asynch::InfallibleOneshotReceiver;
+
+use crate::{
+    client::{connection, PeerInformation},
+    NetworkZone, PeerError, PeerRequest, PeerResponse, SharedError,
+};
+
+/// A weak handle to a [`Client`](super::Client).
+///
+/// When this is dropped the peer will not be disconnected.
+pub struct WeakClient<N: NetworkZone> {
+    /// Information on the connected peer.
+    pub info: PeerInformation<N::Addr>,
+
+    /// The channel to the [`Connection`](connection::Connection) task.
+    pub(super) connection_tx: mpsc::WeakSender<connection::ConnectionTaskRequest>,
+
+    /// The semaphore that limits the requests sent to the peer.
+    pub(super) semaphore: PollSemaphore,
+    /// A permit for the semaphore, will be [`Some`] after `poll_ready` returns ready.
+    pub(super) permit: Option<OwnedSemaphorePermit>,
+
+    /// The error slot shared between the [`Client`] and [`Connection`](connection::Connection).
+    pub(super) error: SharedError<PeerError>,
+}
+
+impl<N: NetworkZone> WeakClient<N> {
+    /// Internal function to set an error on the [`SharedError`].
+    fn set_err(&self, err: PeerError) -> tower::BoxError {
+        let err_str = err.to_string();
+        match self.error.try_insert_err(err) {
+            Ok(()) => err_str,
+            Err(e) => e.to_string(),
+        }
+        .into()
+    }
+}
+
+impl<Z: NetworkZone> Service<PeerRequest> for WeakClient<Z> {
+    type Response = PeerResponse;
+    type Error = tower::BoxError;
+    type Future = InfallibleOneshotReceiver<Result<Self::Response, Self::Error>>;
+
+    fn poll_ready(&mut self, cx: &mut Context<'_>) -> Poll<Result<(), Self::Error>> {
+        if let Some(err) = self.error.try_get_err() {
+            return Poll::Ready(Err(err.to_string().into()));
+        }
+
+        if self.connection_tx.strong_count() == 0 {
+            let err = self.set_err(PeerError::ClientChannelClosed);
+            return Poll::Ready(Err(err));
+        }
+
+        if self.permit.is_some() {
+            return Poll::Ready(Ok(()));
+        }
+
+        let permit = ready!(self.semaphore.poll_acquire(cx))
+            .expect("Client semaphore should not be closed!");
+
+        self.permit = Some(permit);
+
+        Poll::Ready(Ok(()))
+    }
+
+    #[expect(clippy::significant_drop_tightening)]
+    fn call(&mut self, request: PeerRequest) -> Self::Future {
+        let permit = self
+            .permit
+            .take()
+            .expect("poll_ready did not return ready before call to call");
+
+        let (tx, rx) = oneshot::channel();
+        let req = connection::ConnectionTaskRequest {
+            response_channel: tx,
+            request,
+            permit: Some(permit),
+        };
+
+        match self.connection_tx.upgrade() {
+            None => {
+                self.set_err(PeerError::ClientChannelClosed);
+
+                let resp = Err(PeerError::ClientChannelClosed.into());
+                drop(req.response_channel.send(resp));
+            }
+            Some(sender) => {
+                if let Err(e) = sender.try_send(req) {
+                    // The connection task could have closed between a call to `poll_ready` and the call to
+                    // `call`, which means if we don't handle the error here the receiver would panic.
+                    use mpsc::error::TrySendError;
+
+                    match e {
+                        TrySendError::Closed(req) | TrySendError::Full(req) => {
+                            self.set_err(PeerError::ClientChannelClosed);
+
+                            let resp = Err(PeerError::ClientChannelClosed.into());
+                            drop(req.response_channel.send(resp));
+                        }
+                    }
+                }
+            }
+        }
+
+        rx.into()
+    }
+}
diff --git a/p2p/p2p/Cargo.toml b/p2p/p2p/Cargo.toml
index 866fb918..e6ebccb8 100644
--- a/p2p/p2p/Cargo.toml
+++ b/p2p/p2p/Cargo.toml
@@ -20,12 +20,12 @@ monero-serai = { workspace = true, features = ["std"] }
 
 tower = { workspace = true, features = ["buffer"] }
 tokio = { workspace = true, features = ["rt", "rt-multi-thread"] }
-rayon = { workspace = true }
 tokio-util = { workspace = true }
+rayon = { workspace = true }
 tokio-stream = { workspace = true, features = ["sync", "time"] }
 futures = { workspace = true, features = ["std"] }
 pin-project = { workspace = true }
-dashmap = { workspace = true }
+indexmap = { workspace = true, features = ["std"] }
 
 thiserror = { workspace = true }
 bytes = { workspace = true, features = ["std"] }
diff --git a/p2p/p2p/src/block_downloader.rs b/p2p/p2p/src/block_downloader.rs
index 5ebcedc0..ee335c9e 100644
--- a/p2p/p2p/src/block_downloader.rs
+++ b/p2p/p2p/src/block_downloader.rs
@@ -8,7 +8,6 @@
 use std::{
     cmp::{max, min, Reverse},
     collections::{BTreeMap, BinaryHeap},
-    sync::Arc,
     time::Duration,
 };
 
@@ -18,7 +17,7 @@ use tokio::{
     task::JoinSet,
     time::{interval, timeout, MissedTickBehavior},
 };
-use tower::{Service, ServiceExt};
+use tower::{util::BoxCloneService, Service, ServiceExt};
 use tracing::{instrument, Instrument, Span};
 
 use cuprate_async_buffer::{BufferAppender, BufferStream};
@@ -27,11 +26,11 @@ use cuprate_p2p_core::{handles::ConnectionHandle, NetworkZone};
 use cuprate_pruning::PruningSeed;
 
 use crate::{
-    client_pool::{ClientPool, ClientPoolDropGuard},
     constants::{
         BLOCK_DOWNLOADER_REQUEST_TIMEOUT, EMPTY_CHAIN_ENTRIES_BEFORE_TOP_ASSUMED, LONG_BAN,
         MAX_BLOCK_BATCH_LEN, MAX_DOWNLOAD_FAILURES,
     },
+    peer_set::ClientDropGuard,
 };
 
 mod block_queue;
@@ -41,6 +40,7 @@ mod request_chain;
 #[cfg(test)]
 mod tests;
 
+use crate::peer_set::{PeerSetRequest, PeerSetResponse};
 use block_queue::{BlockQueue, ReadyQueueBatch};
 use chain_tracker::{BlocksToRetrieve, ChainEntry, ChainTracker};
 use download_batch::download_batch_task;
@@ -135,7 +135,7 @@ pub enum ChainSvcResponse {
 /// call this function again, so it can start the search again.
 #[instrument(level = "error", skip_all, name = "block_downloader")]
 pub fn download_blocks<N: NetworkZone, C>(
-    client_pool: Arc<ClientPool<N>>,
+    peer_set: BoxCloneService<PeerSetRequest, PeerSetResponse<N>, tower::BoxError>,
     our_chain_svc: C,
     config: BlockDownloaderConfig,
 ) -> BufferStream<BlockBatch>
@@ -147,8 +147,7 @@ where
 {
     let (buffer_appender, buffer_stream) = cuprate_async_buffer::new_buffer(config.buffer_size);
 
-    let block_downloader =
-        BlockDownloader::new(client_pool, our_chain_svc, buffer_appender, config);
+    let block_downloader = BlockDownloader::new(peer_set, our_chain_svc, buffer_appender, config);
 
     tokio::spawn(
         block_downloader
@@ -186,8 +185,8 @@ where
 /// - download an already requested batch of blocks (this might happen due to an error in the previous request
 ///   or because the queue of ready blocks is too large, so we need the oldest block to clear it).
 struct BlockDownloader<N: NetworkZone, C> {
-    /// The client pool.
-    client_pool: Arc<ClientPool<N>>,
+    /// The peer set.
+    peer_set: BoxCloneService<PeerSetRequest, PeerSetResponse<N>, tower::BoxError>,
 
     /// The service that holds our current chain state.
     our_chain_svc: C,
@@ -208,7 +207,7 @@ struct BlockDownloader<N: NetworkZone, C> {
     ///
     /// Returns a result of the chain entry or an error.
     #[expect(clippy::type_complexity)]
-    chain_entry_task: JoinSet<Result<(ClientPoolDropGuard<N>, ChainEntry<N>), BlockDownloadError>>,
+    chain_entry_task: JoinSet<Result<(ClientDropGuard<N>, ChainEntry<N>), BlockDownloadError>>,
 
     /// The current inflight requests.
     ///
@@ -235,13 +234,13 @@ where
 {
     /// Creates a new [`BlockDownloader`]
     fn new(
-        client_pool: Arc<ClientPool<N>>,
+        peer_set: BoxCloneService<PeerSetRequest, PeerSetResponse<N>, tower::BoxError>,
         our_chain_svc: C,
         buffer_appender: BufferAppender<BlockBatch>,
         config: BlockDownloaderConfig,
     ) -> Self {
         Self {
-            client_pool,
+            peer_set,
             our_chain_svc,
             amount_of_blocks_to_request: config.initial_batch_len,
             amount_of_blocks_to_request_updated_at: 0,
@@ -259,7 +258,7 @@ where
     fn check_pending_peers(
         &mut self,
         chain_tracker: &mut ChainTracker<N>,
-        pending_peers: &mut BTreeMap<PruningSeed, Vec<ClientPoolDropGuard<N>>>,
+        pending_peers: &mut BTreeMap<PruningSeed, Vec<ClientDropGuard<N>>>,
     ) {
         tracing::debug!("Checking if we can give any work to pending peers.");
 
@@ -286,11 +285,11 @@ where
     /// This function will find the batch(es) that we are waiting on to clear our ready queue and sends another request
     /// for them.
     ///
-    /// Returns the [`ClientPoolDropGuard`] back if it doesn't have the batch according to its pruning seed.
+    /// Returns the [`ClientDropGuard`] back if it doesn't have the batch according to its pruning seed.
     fn request_inflight_batch_again(
         &mut self,
-        client: ClientPoolDropGuard<N>,
-    ) -> Option<ClientPoolDropGuard<N>> {
+        client: ClientDropGuard<N>,
+    ) -> Option<ClientDropGuard<N>> {
         tracing::debug!(
             "Requesting an inflight batch, current ready queue size: {}",
             self.block_queue.size()
@@ -336,13 +335,13 @@ where
     ///
     /// The batch requested will depend on our current state, failed batches will be prioritised.
     ///
-    /// Returns the [`ClientPoolDropGuard`] back if it doesn't have the data we currently need according
+    /// Returns the [`ClientDropGuard`] back if it doesn't have the data we currently need according
     /// to its pruning seed.
     fn request_block_batch(
         &mut self,
         chain_tracker: &mut ChainTracker<N>,
-        client: ClientPoolDropGuard<N>,
-    ) -> Option<ClientPoolDropGuard<N>> {
+        client: ClientDropGuard<N>,
+    ) -> Option<ClientDropGuard<N>> {
         tracing::trace!("Using peer to request a batch of blocks.");
         // First look to see if we have any failed requests.
         while let Some(failed_request) = self.failed_batches.peek() {
@@ -416,13 +415,13 @@ where
     /// This function will use our current state to decide if we should send a request for a chain entry
     /// or if we should request a batch of blocks.
     ///
-    /// Returns the [`ClientPoolDropGuard`] back if it doesn't have the data we currently need according
+    /// Returns the [`ClientDropGuard`] back if it doesn't have the data we currently need according
     /// to its pruning seed.
     fn try_handle_free_client(
         &mut self,
         chain_tracker: &mut ChainTracker<N>,
-        client: ClientPoolDropGuard<N>,
-    ) -> Option<ClientPoolDropGuard<N>> {
+        client: ClientDropGuard<N>,
+    ) -> Option<ClientDropGuard<N>> {
         // We send 2 requests, so if one of them is slow or doesn't have the next chain, we still have a backup.
         if self.chain_entry_task.len() < 2
             // If we have had too many failures then assume the tip has been found so no more chain entries.
@@ -463,7 +462,7 @@ where
     async fn check_for_free_clients(
         &mut self,
         chain_tracker: &mut ChainTracker<N>,
-        pending_peers: &mut BTreeMap<PruningSeed, Vec<ClientPoolDropGuard<N>>>,
+        pending_peers: &mut BTreeMap<PruningSeed, Vec<ClientDropGuard<N>>>,
     ) -> Result<(), BlockDownloadError> {
         tracing::debug!("Checking for free peers");
 
@@ -478,10 +477,19 @@ where
             panic!("Chain service returned wrong response.");
         };
 
-        for client in self
-            .client_pool
-            .clients_with_more_cumulative_difficulty(current_cumulative_difficulty)
-        {
+        let PeerSetResponse::PeersWithMorePoW(clients) = self
+            .peer_set
+            .ready()
+            .await?
+            .call(PeerSetRequest::PeersWithMorePoW(
+                current_cumulative_difficulty,
+            ))
+            .await?
+        else {
+            unreachable!();
+        };
+
+        for client in clients {
             pending_peers
                 .entry(client.info.pruning_seed)
                 .or_default()
@@ -497,9 +505,9 @@ where
     async fn handle_download_batch_res(
         &mut self,
         start_height: usize,
-        res: Result<(ClientPoolDropGuard<N>, BlockBatch), BlockDownloadError>,
+        res: Result<(ClientDropGuard<N>, BlockBatch), BlockDownloadError>,
         chain_tracker: &mut ChainTracker<N>,
-        pending_peers: &mut BTreeMap<PruningSeed, Vec<ClientPoolDropGuard<N>>>,
+        pending_peers: &mut BTreeMap<PruningSeed, Vec<ClientDropGuard<N>>>,
     ) -> Result<(), BlockDownloadError> {
         tracing::debug!("Handling block download response");
 
@@ -593,7 +601,7 @@ where
     /// Starts the main loop of the block downloader.
     async fn run(mut self) -> Result<(), BlockDownloadError> {
         let mut chain_tracker =
-            initial_chain_search(&self.client_pool, &mut self.our_chain_svc).await?;
+            initial_chain_search(&mut self.peer_set, &mut self.our_chain_svc).await?;
 
         let mut pending_peers = BTreeMap::new();
 
@@ -662,7 +670,7 @@ struct BlockDownloadTaskResponse<N: NetworkZone> {
     /// The start height of the batch.
     start_height: usize,
     /// A result containing the batch or an error.
-    result: Result<(ClientPoolDropGuard<N>, BlockBatch), BlockDownloadError>,
+    result: Result<(ClientDropGuard<N>, BlockBatch), BlockDownloadError>,
 }
 
 /// Returns if a peer has all the blocks in a range, according to its [`PruningSeed`].
diff --git a/p2p/p2p/src/block_downloader/download_batch.rs b/p2p/p2p/src/block_downloader/download_batch.rs
index bbb14b3b..ef621ce8 100644
--- a/p2p/p2p/src/block_downloader/download_batch.rs
+++ b/p2p/p2p/src/block_downloader/download_batch.rs
@@ -16,8 +16,8 @@ use cuprate_wire::protocol::{GetObjectsRequest, GetObjectsResponse};
 
 use crate::{
     block_downloader::{BlockBatch, BlockDownloadError, BlockDownloadTaskResponse},
-    client_pool::ClientPoolDropGuard,
     constants::{BLOCK_DOWNLOADER_REQUEST_TIMEOUT, MAX_TRANSACTION_BLOB_SIZE, MEDIUM_BAN},
+    peer_set::ClientDropGuard,
 };
 
 /// Attempts to request a batch of blocks from a peer, returning [`BlockDownloadTaskResponse`].
@@ -32,7 +32,7 @@ use crate::{
 )]
 #[expect(clippy::used_underscore_binding)]
 pub async fn download_batch_task<N: NetworkZone>(
-    client: ClientPoolDropGuard<N>,
+    client: ClientDropGuard<N>,
     ids: ByteArrayVec<32>,
     previous_id: [u8; 32],
     expected_start_height: usize,
@@ -49,11 +49,11 @@ pub async fn download_batch_task<N: NetworkZone>(
 /// This function will validate the blocks that were downloaded were the ones asked for and that they match
 /// the expected height.
 async fn request_batch_from_peer<N: NetworkZone>(
-    mut client: ClientPoolDropGuard<N>,
+    mut client: ClientDropGuard<N>,
     ids: ByteArrayVec<32>,
     previous_id: [u8; 32],
     expected_start_height: usize,
-) -> Result<(ClientPoolDropGuard<N>, BlockBatch), BlockDownloadError> {
+) -> Result<(ClientDropGuard<N>, BlockBatch), BlockDownloadError> {
     let request = PeerRequest::Protocol(ProtocolRequest::GetObjects(GetObjectsRequest {
         blocks: ids.clone(),
         pruned: false,
diff --git a/p2p/p2p/src/block_downloader/request_chain.rs b/p2p/p2p/src/block_downloader/request_chain.rs
index d6a2a0af..4e0f855b 100644
--- a/p2p/p2p/src/block_downloader/request_chain.rs
+++ b/p2p/p2p/src/block_downloader/request_chain.rs
@@ -1,7 +1,7 @@
-use std::{mem, sync::Arc};
+use std::mem;
 
 use tokio::{task::JoinSet, time::timeout};
-use tower::{Service, ServiceExt};
+use tower::{util::BoxCloneService, Service, ServiceExt};
 use tracing::{instrument, Instrument, Span};
 
 use cuprate_p2p_core::{
@@ -15,11 +15,11 @@ use crate::{
         chain_tracker::{ChainEntry, ChainTracker},
         BlockDownloadError, ChainSvcRequest, ChainSvcResponse,
     },
-    client_pool::{ClientPool, ClientPoolDropGuard},
     constants::{
         BLOCK_DOWNLOADER_REQUEST_TIMEOUT, INITIAL_CHAIN_REQUESTS_TO_SEND,
         MAX_BLOCKS_IDS_IN_CHAIN_ENTRY, MEDIUM_BAN,
     },
+    peer_set::{ClientDropGuard, PeerSetRequest, PeerSetResponse},
 };
 
 /// Request a chain entry from a peer.
@@ -27,9 +27,9 @@ use crate::{
 /// Because the block downloader only follows and downloads one chain we only have to send the block hash of
 /// top block we have found and the genesis block, this is then called `short_history`.
 pub(crate) async fn request_chain_entry_from_peer<N: NetworkZone>(
-    mut client: ClientPoolDropGuard<N>,
+    mut client: ClientDropGuard<N>,
     short_history: [[u8; 32]; 2],
-) -> Result<(ClientPoolDropGuard<N>, ChainEntry<N>), BlockDownloadError> {
+) -> Result<(ClientDropGuard<N>, ChainEntry<N>), BlockDownloadError> {
     let PeerResponse::Protocol(ProtocolResponse::GetChain(chain_res)) = client
         .ready()
         .await?
@@ -80,7 +80,7 @@ pub(crate) async fn request_chain_entry_from_peer<N: NetworkZone>(
 /// We then wait for their response and choose the peer who claims the highest cumulative difficulty.
 #[instrument(level = "error", skip_all)]
 pub async fn initial_chain_search<N: NetworkZone, C>(
-    client_pool: &Arc<ClientPool<N>>,
+    peer_set: &mut BoxCloneService<PeerSetRequest, PeerSetResponse<N>, tower::BoxError>,
     mut our_chain_svc: C,
 ) -> Result<ChainTracker<N>, BlockDownloadError>
 where
@@ -102,9 +102,15 @@ where
 
     let our_genesis = *block_ids.last().expect("Blockchain had no genesis block.");
 
-    let mut peers = client_pool
-        .clients_with_more_cumulative_difficulty(cumulative_difficulty)
-        .into_iter();
+    let PeerSetResponse::PeersWithMorePoW(clients) = peer_set
+        .ready()
+        .await?
+        .call(PeerSetRequest::PeersWithMorePoW(cumulative_difficulty))
+        .await?
+    else {
+        unreachable!();
+    };
+    let mut peers = clients.into_iter();
 
     let mut futs = JoinSet::new();
 
diff --git a/p2p/p2p/src/block_downloader/tests.rs b/p2p/p2p/src/block_downloader/tests.rs
index dd07cce9..2d00358e 100644
--- a/p2p/p2p/src/block_downloader/tests.rs
+++ b/p2p/p2p/src/block_downloader/tests.rs
@@ -14,8 +14,8 @@ use monero_serai::{
     transaction::{Input, Timelock, Transaction, TransactionPrefix},
 };
 use proptest::{collection::vec, prelude::*};
-use tokio::time::timeout;
-use tower::{service_fn, Service};
+use tokio::{sync::mpsc, time::timeout};
+use tower::{buffer::Buffer, service_fn, Service, ServiceExt};
 
 use cuprate_fixed_bytes::ByteArrayVec;
 use cuprate_p2p_core::{
@@ -31,7 +31,7 @@ use cuprate_wire::{
 
 use crate::{
     block_downloader::{download_blocks, BlockDownloaderConfig, ChainSvcRequest, ChainSvcResponse},
-    client_pool::ClientPool,
+    peer_set::PeerSet,
 };
 
 proptest! {
@@ -48,19 +48,20 @@ proptest! {
 
         let tokio_pool = tokio::runtime::Builder::new_multi_thread().enable_all().build().unwrap();
 
-        #[expect(clippy::significant_drop_tightening)]
         tokio_pool.block_on(async move {
             timeout(Duration::from_secs(600), async move {
-                let client_pool = ClientPool::new();
+                let (new_connection_tx, new_connection_rx) = mpsc::channel(peers);
+
+                let peer_set = PeerSet::new(new_connection_rx);
 
                 for _ in 0..peers {
                     let client = mock_block_downloader_client(Arc::clone(&blockchain));
 
-                    client_pool.add_new_client(client);
+                    new_connection_tx.try_send(client).unwrap();
                 }
 
                 let stream = download_blocks(
-                    client_pool,
+                    Buffer::new(peer_set, 10).boxed_clone(),
                     OurChainSvc {
                         genesis: *blockchain.blocks.first().unwrap().0
                     },
diff --git a/p2p/p2p/src/client_pool.rs b/p2p/p2p/src/client_pool.rs
deleted file mode 100644
index 67c8f112..00000000
--- a/p2p/p2p/src/client_pool.rs
+++ /dev/null
@@ -1,188 +0,0 @@
-//! # Client Pool.
-//!
-//! The [`ClientPool`], is a pool of currently connected peers that can be pulled from.
-//! It does _not_ necessarily contain every connected peer as another place could have
-//! taken a peer from the pool.
-//!
-//! When taking peers from the pool they are wrapped in [`ClientPoolDropGuard`], which
-//! returns the peer to the pool when it is dropped.
-//!
-//! Internally the pool is a [`DashMap`] which means care should be taken in `async` code
-//! as internally this uses blocking `RwLock`s.
-use std::sync::Arc;
-
-use dashmap::DashMap;
-use tokio::sync::mpsc;
-use tracing::{Instrument, Span};
-
-use cuprate_p2p_core::{
-    client::{Client, InternalPeerID},
-    handles::ConnectionHandle,
-    ConnectionDirection, NetworkZone,
-};
-
-pub(crate) mod disconnect_monitor;
-mod drop_guard_client;
-
-pub use drop_guard_client::ClientPoolDropGuard;
-
-/// The client pool, which holds currently connected free peers.
-///
-/// See the [module docs](self) for more.
-pub struct ClientPool<N: NetworkZone> {
-    /// The connected [`Client`]s.
-    clients: DashMap<InternalPeerID<N::Addr>, Client<N>>,
-    /// A channel to send new peer ids down to monitor for disconnect.
-    new_connection_tx: mpsc::UnboundedSender<(ConnectionHandle, InternalPeerID<N::Addr>)>,
-}
-
-impl<N: NetworkZone> ClientPool<N> {
-    /// Returns a new [`ClientPool`] wrapped in an [`Arc`].
-    pub fn new() -> Arc<Self> {
-        let (tx, rx) = mpsc::unbounded_channel();
-
-        let pool = Arc::new(Self {
-            clients: DashMap::new(),
-            new_connection_tx: tx,
-        });
-
-        tokio::spawn(
-            disconnect_monitor::disconnect_monitor(rx, Arc::clone(&pool))
-                .instrument(Span::current()),
-        );
-
-        pool
-    }
-
-    /// Adds a [`Client`] to the pool, the client must have previously been taken from the
-    /// pool.
-    ///
-    /// See [`ClientPool::add_new_client`] to add a [`Client`] which was not taken from the pool before.
-    ///
-    /// # Panics
-    /// This function panics if `client` already exists in the pool.
-    fn add_client(&self, client: Client<N>) {
-        let handle = client.info.handle.clone();
-        let id = client.info.id;
-
-        // Fast path: if the client is disconnected don't add it to the peer set.
-        if handle.is_closed() {
-            return;
-        }
-
-        assert!(self.clients.insert(id, client).is_none());
-
-        // We have to check this again otherwise we could have a race condition where a
-        // peer is disconnected after the first check, the disconnect monitor tries to remove it,
-        // and then it is added to the pool.
-        if handle.is_closed() {
-            self.remove_client(&id);
-        }
-    }
-
-    /// Adds a _new_ [`Client`] to the pool, this client should be a new connection, and not already
-    /// from the pool.
-    ///
-    /// # Panics
-    /// This function panics if `client` already exists in the pool.
-    pub fn add_new_client(&self, client: Client<N>) {
-        self.new_connection_tx
-            .send((client.info.handle.clone(), client.info.id))
-            .unwrap();
-
-        self.add_client(client);
-    }
-
-    /// Remove a [`Client`] from the pool.
-    ///
-    /// [`None`] is returned if the client did not exist in the pool.
-    fn remove_client(&self, peer: &InternalPeerID<N::Addr>) -> Option<Client<N>> {
-        self.clients.remove(peer).map(|(_, client)| client)
-    }
-
-    /// Borrows a [`Client`] from the pool.
-    ///
-    /// The [`Client`] is wrapped in [`ClientPoolDropGuard`] which
-    /// will return the client to the pool when it's dropped.
-    ///
-    /// See [`Self::borrow_clients`] for borrowing multiple clients.
-    pub fn borrow_client(
-        self: &Arc<Self>,
-        peer: &InternalPeerID<N::Addr>,
-    ) -> Option<ClientPoolDropGuard<N>> {
-        self.remove_client(peer).map(|client| ClientPoolDropGuard {
-            pool: Arc::clone(self),
-            client: Some(client),
-        })
-    }
-
-    /// Borrows multiple [`Client`]s from the pool.
-    ///
-    /// Note that the returned iterator is not guaranteed to contain every peer asked for.
-    ///
-    /// See [`Self::borrow_client`] for borrowing a single client.
-    pub fn borrow_clients<'a, 'b>(
-        self: &'a Arc<Self>,
-        peers: &'b [InternalPeerID<N::Addr>],
-    ) -> impl Iterator<Item = ClientPoolDropGuard<N>> + sealed::Captures<(&'a (), &'b ())> {
-        peers.iter().filter_map(|peer| self.borrow_client(peer))
-    }
-
-    /// Borrows all [`Client`]s from the pool that have claimed a higher cumulative difficulty than
-    /// the amount passed in.
-    ///
-    /// The [`Client`]s are wrapped in [`ClientPoolDropGuard`] which
-    /// will return the clients to the pool when they are dropped.
-    pub fn clients_with_more_cumulative_difficulty(
-        self: &Arc<Self>,
-        cumulative_difficulty: u128,
-    ) -> Vec<ClientPoolDropGuard<N>> {
-        let peers = self
-            .clients
-            .iter()
-            .filter_map(|element| {
-                let peer_sync_info = element.value().info.core_sync_data.lock().unwrap();
-
-                if peer_sync_info.cumulative_difficulty() > cumulative_difficulty {
-                    Some(*element.key())
-                } else {
-                    None
-                }
-            })
-            .collect::<Vec<_>>();
-
-        self.borrow_clients(&peers).collect()
-    }
-
-    /// Checks all clients in the pool checking if any claim a higher cumulative difficulty than the
-    /// amount specified.
-    pub fn contains_client_with_more_cumulative_difficulty(
-        &self,
-        cumulative_difficulty: u128,
-    ) -> bool {
-        self.clients.iter().any(|element| {
-            let sync_data = element.value().info.core_sync_data.lock().unwrap();
-            sync_data.cumulative_difficulty() > cumulative_difficulty
-        })
-    }
-
-    /// Returns the first outbound peer when iterating over the peers.
-    pub fn outbound_client(self: &Arc<Self>) -> Option<ClientPoolDropGuard<N>> {
-        let client = self
-            .clients
-            .iter()
-            .find(|element| element.value().info.direction == ConnectionDirection::Outbound)?;
-        let id = *client.key();
-
-        Some(self.borrow_client(&id).unwrap())
-    }
-}
-
-mod sealed {
-    /// TODO: Remove me when 2024 Rust
-    ///
-    /// <https://rust-lang.github.io/rfcs/3498-lifetime-capture-rules-2024.html#the-captures-trick>
-    pub trait Captures<U> {}
-
-    impl<T: ?Sized, U> Captures<U> for T {}
-}
diff --git a/p2p/p2p/src/client_pool/disconnect_monitor.rs b/p2p/p2p/src/client_pool/disconnect_monitor.rs
deleted file mode 100644
index f54b5606..00000000
--- a/p2p/p2p/src/client_pool/disconnect_monitor.rs
+++ /dev/null
@@ -1,83 +0,0 @@
-//! # Disconnect Monitor
-//!
-//! This module contains the [`disconnect_monitor`] task, which monitors connected peers for disconnection
-//! and then removes them from the [`ClientPool`] if they do.
-use std::{
-    future::Future,
-    pin::Pin,
-    sync::Arc,
-    task::{Context, Poll},
-};
-
-use futures::{stream::FuturesUnordered, StreamExt};
-use tokio::sync::mpsc;
-use tokio_util::sync::WaitForCancellationFutureOwned;
-use tracing::instrument;
-
-use cuprate_p2p_core::{client::InternalPeerID, handles::ConnectionHandle, NetworkZone};
-
-use super::ClientPool;
-
-/// The disconnect monitor task.
-#[instrument(level = "info", skip_all)]
-pub async fn disconnect_monitor<N: NetworkZone>(
-    mut new_connection_rx: mpsc::UnboundedReceiver<(ConnectionHandle, InternalPeerID<N::Addr>)>,
-    client_pool: Arc<ClientPool<N>>,
-) {
-    // We need to hold a weak reference otherwise the client pool and this would hold a reference to
-    // each other causing the pool to be leaked.
-    let weak_client_pool = Arc::downgrade(&client_pool);
-    drop(client_pool);
-
-    tracing::info!("Starting peer disconnect monitor.");
-
-    let mut futs: FuturesUnordered<PeerDisconnectFut<N>> = FuturesUnordered::new();
-
-    loop {
-        tokio::select! {
-            Some((con_handle, peer_id)) = new_connection_rx.recv() => {
-                tracing::debug!("Monitoring {peer_id} for disconnect");
-                futs.push(PeerDisconnectFut {
-                    closed_fut: con_handle.closed(),
-                    peer_id: Some(peer_id),
-                });
-            }
-            Some(peer_id) = futs.next() => {
-                tracing::debug!("{peer_id} has disconnected, removing from client pool.");
-                let Some(pool) = weak_client_pool.upgrade() else {
-                    tracing::info!("Peer disconnect monitor shutting down.");
-                    return;
-                };
-
-                pool.remove_client(&peer_id);
-                drop(pool);
-            }
-            else => {
-                tracing::info!("Peer disconnect monitor shutting down.");
-                return;
-            }
-        }
-    }
-}
-
-/// A [`Future`] that resolves when a peer disconnects.
-#[pin_project::pin_project]
-pub(crate) struct PeerDisconnectFut<N: NetworkZone> {
-    /// The inner [`Future`] that resolves when a peer disconnects.
-    #[pin]
-    pub(crate) closed_fut: WaitForCancellationFutureOwned,
-    /// The peers ID.
-    pub(crate) peer_id: Option<InternalPeerID<N::Addr>>,
-}
-
-impl<N: NetworkZone> Future for PeerDisconnectFut<N> {
-    type Output = InternalPeerID<N::Addr>;
-
-    fn poll(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
-        let this = self.project();
-
-        this.closed_fut
-            .poll(cx)
-            .map(|()| this.peer_id.take().unwrap())
-    }
-}
diff --git a/p2p/p2p/src/client_pool/drop_guard_client.rs b/p2p/p2p/src/client_pool/drop_guard_client.rs
deleted file mode 100644
index b10c4e9c..00000000
--- a/p2p/p2p/src/client_pool/drop_guard_client.rs
+++ /dev/null
@@ -1,41 +0,0 @@
-use std::{
-    ops::{Deref, DerefMut},
-    sync::Arc,
-};
-
-use cuprate_p2p_core::{client::Client, NetworkZone};
-
-use crate::client_pool::ClientPool;
-
-/// A wrapper around [`Client`] which returns the client to the [`ClientPool`] when dropped.
-pub struct ClientPoolDropGuard<N: NetworkZone> {
-    /// The [`ClientPool`] to return the peer to.
-    pub(super) pool: Arc<ClientPool<N>>,
-    /// The [`Client`].
-    ///
-    /// This is set to [`Some`] when this guard is created, then
-    /// [`take`](Option::take)n and returned to the pool when dropped.
-    pub(super) client: Option<Client<N>>,
-}
-
-impl<N: NetworkZone> Deref for ClientPoolDropGuard<N> {
-    type Target = Client<N>;
-
-    fn deref(&self) -> &Self::Target {
-        self.client.as_ref().unwrap()
-    }
-}
-
-impl<N: NetworkZone> DerefMut for ClientPoolDropGuard<N> {
-    fn deref_mut(&mut self) -> &mut Self::Target {
-        self.client.as_mut().unwrap()
-    }
-}
-
-impl<N: NetworkZone> Drop for ClientPoolDropGuard<N> {
-    fn drop(&mut self) {
-        let client = self.client.take().unwrap();
-
-        self.pool.add_client(client);
-    }
-}
diff --git a/p2p/p2p/src/connection_maintainer.rs b/p2p/p2p/src/connection_maintainer.rs
index cd9d9311..245fbf12 100644
--- a/p2p/p2p/src/connection_maintainer.rs
+++ b/p2p/p2p/src/connection_maintainer.rs
@@ -21,7 +21,6 @@ use cuprate_p2p_core::{
 };
 
 use crate::{
-    client_pool::ClientPool,
     config::P2PConfig,
     constants::{HANDSHAKE_TIMEOUT, MAX_SEED_CONNECTIONS, OUTBOUND_CONNECTION_ATTEMPT_TIMEOUT},
 };
@@ -46,7 +45,7 @@ pub struct MakeConnectionRequest {
 /// This handles maintaining a minimum number of connections and making extra connections when needed, upto a maximum.
 pub struct OutboundConnectionKeeper<N: NetworkZone, A, C> {
     /// The pool of currently connected peers.
-    pub client_pool: Arc<ClientPool<N>>,
+    pub new_peers_tx: mpsc::Sender<Client<N>>,
     /// The channel that tells us to make new _extra_ outbound connections.
     pub make_connection_rx: mpsc::Receiver<MakeConnectionRequest>,
     /// The address book service
@@ -77,7 +76,7 @@ where
 {
     pub fn new(
         config: P2PConfig<N>,
-        client_pool: Arc<ClientPool<N>>,
+        new_peers_tx: mpsc::Sender<Client<N>>,
         make_connection_rx: mpsc::Receiver<MakeConnectionRequest>,
         address_book_svc: A,
         connector_svc: C,
@@ -86,7 +85,7 @@ where
             .expect("Gray peer percent is incorrect should be 0..=1");
 
         Self {
-            client_pool,
+            new_peers_tx,
             make_connection_rx,
             address_book_svc,
             connector_svc,
@@ -149,7 +148,7 @@ where
     /// Connects to a given outbound peer.
     #[instrument(level = "info", skip_all)]
     async fn connect_to_outbound_peer(&mut self, permit: OwnedSemaphorePermit, addr: N::Addr) {
-        let client_pool = Arc::clone(&self.client_pool);
+        let new_peers_tx = self.new_peers_tx.clone();
         let connection_fut = self
             .connector_svc
             .ready()
@@ -164,7 +163,7 @@ where
             async move {
                 #[expect(clippy::significant_drop_in_scrutinee)]
                 if let Ok(Ok(peer)) = timeout(HANDSHAKE_TIMEOUT, connection_fut).await {
-                    client_pool.add_new_client(peer);
+                    drop(new_peers_tx.send(peer).await);
                 }
             }
             .instrument(Span::current()),
diff --git a/p2p/p2p/src/inbound_server.rs b/p2p/p2p/src/inbound_server.rs
index 6e793bd1..0479560b 100644
--- a/p2p/p2p/src/inbound_server.rs
+++ b/p2p/p2p/src/inbound_server.rs
@@ -6,7 +6,7 @@ use std::{pin::pin, sync::Arc};
 
 use futures::{SinkExt, StreamExt};
 use tokio::{
-    sync::Semaphore,
+    sync::{mpsc, Semaphore},
     task::JoinSet,
     time::{sleep, timeout},
 };
@@ -24,7 +24,6 @@ use cuprate_wire::{
 };
 
 use crate::{
-    client_pool::ClientPool,
     constants::{
         HANDSHAKE_TIMEOUT, INBOUND_CONNECTION_COOL_DOWN, PING_REQUEST_CONCURRENCY,
         PING_REQUEST_TIMEOUT,
@@ -36,7 +35,7 @@ use crate::{
 /// and initiate handshake if needed, after verifying the address isn't banned.
 #[instrument(level = "warn", skip_all)]
 pub async fn inbound_server<N, HS, A>(
-    client_pool: Arc<ClientPool<N>>,
+    new_connection_tx: mpsc::Sender<Client<N>>,
     mut handshaker: HS,
     mut address_book: A,
     config: P2PConfig<N>,
@@ -111,13 +110,13 @@ where
                 permit: Some(permit),
             });
 
-            let cloned_pool = Arc::clone(&client_pool);
+            let new_connection_tx = new_connection_tx.clone();
 
             tokio::spawn(
                 async move {
                     let client = timeout(HANDSHAKE_TIMEOUT, fut).await;
                     if let Ok(Ok(peer)) = client {
-                        cloned_pool.add_new_client(peer);
+                        drop(new_connection_tx.send(peer).await);
                     }
                 }
                 .instrument(Span::current()),
diff --git a/p2p/p2p/src/lib.rs b/p2p/p2p/src/lib.rs
index 541784c9..fb506582 100644
--- a/p2p/p2p/src/lib.rs
+++ b/p2p/p2p/src/lib.rs
@@ -18,17 +18,18 @@ use cuprate_p2p_core::{
 
 pub mod block_downloader;
 mod broadcast;
-pub mod client_pool;
 pub mod config;
 pub mod connection_maintainer;
 pub mod constants;
 mod inbound_server;
+mod peer_set;
 
 use block_downloader::{BlockBatch, BlockDownloaderConfig, ChainSvcRequest, ChainSvcResponse};
 pub use broadcast::{BroadcastRequest, BroadcastSvc};
-pub use client_pool::{ClientPool, ClientPoolDropGuard};
 pub use config::{AddressBookConfig, P2PConfig};
 use connection_maintainer::MakeConnectionRequest;
+use peer_set::PeerSet;
+pub use peer_set::{ClientDropGuard, PeerSetRequest, PeerSetResponse};
 
 /// Initializes the P2P [`NetworkInterface`] for a specific [`NetworkZone`].
 ///
@@ -54,7 +55,10 @@ where
         cuprate_address_book::init_address_book(config.address_book_config.clone()).await?;
     let address_book = Buffer::new(
         address_book,
-        config.max_inbound_connections + config.outbound_connections,
+        config
+            .max_inbound_connections
+            .checked_add(config.outbound_connections)
+            .unwrap(),
     );
 
     // Use the default config. Changing the defaults affects tx fluff times, which could affect D++ so for now don't allow changing
@@ -83,19 +87,25 @@ where
 
     let outbound_handshaker = outbound_handshaker_builder.build();
 
-    let client_pool = ClientPool::new();
-
+    let (new_connection_tx, new_connection_rx) = mpsc::channel(
+        config
+            .outbound_connections
+            .checked_add(config.max_inbound_connections)
+            .unwrap(),
+    );
     let (make_connection_tx, make_connection_rx) = mpsc::channel(3);
 
     let outbound_connector = Connector::new(outbound_handshaker);
     let outbound_connection_maintainer = connection_maintainer::OutboundConnectionKeeper::new(
         config.clone(),
-        Arc::clone(&client_pool),
+        new_connection_tx.clone(),
         make_connection_rx,
         address_book.clone(),
         outbound_connector,
     );
 
+    let peer_set = PeerSet::new(new_connection_rx);
+
     let mut background_tasks = JoinSet::new();
 
     background_tasks.spawn(
@@ -105,7 +115,7 @@ where
     );
     background_tasks.spawn(
         inbound_server::inbound_server(
-            Arc::clone(&client_pool),
+            new_connection_tx,
             inbound_handshaker,
             address_book.clone(),
             config,
@@ -121,7 +131,7 @@ where
     );
 
     Ok(NetworkInterface {
-        pool: client_pool,
+        peer_set: Buffer::new(peer_set, 10).boxed_clone(),
         broadcast_svc,
         make_connection_tx,
         address_book: address_book.boxed_clone(),
@@ -133,7 +143,7 @@ where
 #[derive(Clone)]
 pub struct NetworkInterface<N: NetworkZone> {
     /// A pool of free connected peers.
-    pool: Arc<ClientPool<N>>,
+    peer_set: BoxCloneService<PeerSetRequest, PeerSetResponse<N>, tower::BoxError>,
     /// A [`Service`] that allows broadcasting to all connected peers.
     broadcast_svc: BroadcastSvc<N>,
     /// A channel to request extra connections.
@@ -163,7 +173,7 @@ impl<N: NetworkZone> NetworkInterface<N> {
             + 'static,
         C::Future: Send + 'static,
     {
-        block_downloader::download_blocks(Arc::clone(&self.pool), our_chain_service, config)
+        block_downloader::download_blocks(self.peer_set.clone(), our_chain_service, config)
     }
 
     /// Returns the address book service.
@@ -173,8 +183,10 @@ impl<N: NetworkZone> NetworkInterface<N> {
         self.address_book.clone()
     }
 
-    /// Borrows the `ClientPool`, for access to connected peers.
-    pub const fn client_pool(&self) -> &Arc<ClientPool<N>> {
-        &self.pool
+    /// Borrows the `PeerSet`, for access to connected peers.
+    pub fn peer_set(
+        &mut self,
+    ) -> &mut BoxCloneService<PeerSetRequest, PeerSetResponse<N>, tower::BoxError> {
+        &mut self.peer_set
     }
 }
diff --git a/p2p/p2p/src/peer_set.rs b/p2p/p2p/src/peer_set.rs
new file mode 100644
index 00000000..498eaafc
--- /dev/null
+++ b/p2p/p2p/src/peer_set.rs
@@ -0,0 +1,217 @@
+use std::{
+    future::{ready, Future, Ready},
+    pin::{pin, Pin},
+    task::{Context, Poll},
+};
+
+use futures::{stream::FuturesUnordered, StreamExt};
+use indexmap::{IndexMap, IndexSet};
+use rand::{seq::index::sample, thread_rng};
+use tokio::sync::mpsc::Receiver;
+use tokio_util::sync::WaitForCancellationFutureOwned;
+use tower::Service;
+
+use cuprate_helper::cast::u64_to_usize;
+use cuprate_p2p_core::{
+    client::{Client, InternalPeerID},
+    ConnectionDirection, NetworkZone,
+};
+
+mod client_wrappers;
+
+pub use client_wrappers::ClientDropGuard;
+use client_wrappers::StoredClient;
+
+/// A request to the peer-set.
+pub enum PeerSetRequest {
+    /// The most claimed proof-of-work from a peer in the peer-set.
+    MostPoWSeen,
+    /// Peers with more cumulative difficulty than the given cumulative difficulty.
+    ///
+    /// Returned peers will be remembered and won't be returned from subsequent calls until the guard is dropped.
+    PeersWithMorePoW(u128),
+    /// A random outbound peer.
+    ///
+    /// The returned peer will be remembered and won't be returned from subsequent calls until the guard is dropped.
+    StemPeer,
+}
+
+/// A response from the peer-set.
+pub enum PeerSetResponse<N: NetworkZone> {
+    /// [`PeerSetRequest::MostPoWSeen`]
+    MostPoWSeen {
+        /// The cumulative difficulty claimed.
+        cumulative_difficulty: u128,
+        /// The height claimed.
+        height: usize,
+        /// The claimed hash of the top block.
+        top_hash: [u8; 32],
+    },
+    /// [`PeerSetRequest::PeersWithMorePoW`]
+    ///
+    /// Returned peers will be remembered and won't be returned from subsequent calls until the guard is dropped.
+    PeersWithMorePoW(Vec<ClientDropGuard<N>>),
+    /// [`PeerSetRequest::StemPeer`]
+    ///
+    /// The returned peer will be remembered and won't be returned from subsequent calls until the guard is dropped.
+    StemPeer(Option<ClientDropGuard<N>>),
+}
+
+/// A [`Future`] that completes when a peer disconnects.
+#[pin_project::pin_project]
+struct ClosedConnectionFuture<N: NetworkZone> {
+    #[pin]
+    fut: WaitForCancellationFutureOwned,
+    id: Option<InternalPeerID<N::Addr>>,
+}
+
+impl<N: NetworkZone> Future for ClosedConnectionFuture<N> {
+    type Output = InternalPeerID<N::Addr>;
+    fn poll(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
+        let this = self.project();
+
+        this.fut.poll(cx).map(|()| this.id.take().unwrap())
+    }
+}
+
+/// A collection of all connected peers on a [`NetworkZone`].
+pub(crate) struct PeerSet<N: NetworkZone> {
+    /// The connected peers.
+    peers: IndexMap<InternalPeerID<N::Addr>, StoredClient<N>>,
+    /// A [`FuturesUnordered`] that resolves when a peer disconnects.
+    closed_connections: FuturesUnordered<ClosedConnectionFuture<N>>,
+    /// The [`InternalPeerID`]s of all outbound peers.
+    outbound_peers: IndexSet<InternalPeerID<N::Addr>>,
+    /// A channel of new peers from the inbound server or outbound connector.
+    new_peers: Receiver<Client<N>>,
+}
+
+impl<N: NetworkZone> PeerSet<N> {
+    pub(crate) fn new(new_peers: Receiver<Client<N>>) -> Self {
+        Self {
+            peers: IndexMap::new(),
+            closed_connections: FuturesUnordered::new(),
+            outbound_peers: IndexSet::new(),
+            new_peers,
+        }
+    }
+
+    /// Polls the new peers channel for newly connected peers.
+    fn poll_new_peers(&mut self, cx: &mut Context<'_>) {
+        while let Poll::Ready(Some(new_peer)) = self.new_peers.poll_recv(cx) {
+            if new_peer.info.direction == ConnectionDirection::Outbound {
+                self.outbound_peers.insert(new_peer.info.id);
+            }
+
+            self.closed_connections.push(ClosedConnectionFuture {
+                fut: new_peer.info.handle.closed(),
+                id: Some(new_peer.info.id),
+            });
+
+            self.peers
+                .insert(new_peer.info.id, StoredClient::new(new_peer));
+        }
+    }
+
+    /// Remove disconnected peers from the peer set.
+    fn remove_dead_peers(&mut self, cx: &mut Context<'_>) {
+        while let Poll::Ready(Some(dead_peer)) = self.closed_connections.poll_next_unpin(cx) {
+            let Some(peer) = self.peers.swap_remove(&dead_peer) else {
+                continue;
+            };
+
+            if peer.client.info.direction == ConnectionDirection::Outbound {
+                self.outbound_peers.swap_remove(&peer.client.info.id);
+            }
+
+            self.peers.swap_remove(&dead_peer);
+        }
+    }
+
+    /// [`PeerSetRequest::MostPoWSeen`]
+    fn most_pow_seen(&self) -> PeerSetResponse<N> {
+        let most_pow_chain = self
+            .peers
+            .values()
+            .map(|peer| {
+                let core_sync_data = peer.client.info.core_sync_data.lock().unwrap();
+
+                (
+                    core_sync_data.cumulative_difficulty(),
+                    u64_to_usize(core_sync_data.current_height),
+                    core_sync_data.top_id,
+                )
+            })
+            .max_by_key(|(cumulative_difficulty, ..)| *cumulative_difficulty)
+            .unwrap_or_default();
+
+        PeerSetResponse::MostPoWSeen {
+            cumulative_difficulty: most_pow_chain.0,
+            height: most_pow_chain.1,
+            top_hash: most_pow_chain.2,
+        }
+    }
+
+    /// [`PeerSetRequest::PeersWithMorePoW`]
+    fn peers_with_more_pow(&self, cumulative_difficulty: u128) -> PeerSetResponse<N> {
+        PeerSetResponse::PeersWithMorePoW(
+            self.peers
+                .values()
+                .filter(|&client| {
+                    !client.is_downloading_blocks()
+                        && client
+                            .client
+                            .info
+                            .core_sync_data
+                            .lock()
+                            .unwrap()
+                            .cumulative_difficulty()
+                            > cumulative_difficulty
+                })
+                .map(StoredClient::downloading_blocks_guard)
+                .collect(),
+        )
+    }
+
+    /// [`PeerSetRequest::StemPeer`]
+    fn random_peer_for_stem(&self) -> PeerSetResponse<N> {
+        PeerSetResponse::StemPeer(
+            sample(
+                &mut thread_rng(),
+                self.outbound_peers.len(),
+                self.outbound_peers.len(),
+            )
+            .into_iter()
+            .find_map(|i| {
+                let peer = self.outbound_peers.get_index(i).unwrap();
+                let client = self.peers.get(peer).unwrap();
+                (!client.is_a_stem_peer()).then(|| client.stem_peer_guard())
+            }),
+        )
+    }
+}
+
+impl<N: NetworkZone> Service<PeerSetRequest> for PeerSet<N> {
+    type Response = PeerSetResponse<N>;
+    type Error = tower::BoxError;
+    type Future = Ready<Result<Self::Response, Self::Error>>;
+
+    fn poll_ready(&mut self, cx: &mut Context<'_>) -> Poll<Result<(), Self::Error>> {
+        self.poll_new_peers(cx);
+        self.remove_dead_peers(cx);
+
+        // TODO: should we return `Pending` if we don't have any peers?
+
+        Poll::Ready(Ok(()))
+    }
+
+    fn call(&mut self, req: PeerSetRequest) -> Self::Future {
+        ready(match req {
+            PeerSetRequest::MostPoWSeen => Ok(self.most_pow_seen()),
+            PeerSetRequest::PeersWithMorePoW(cumulative_difficulty) => {
+                Ok(self.peers_with_more_pow(cumulative_difficulty))
+            }
+            PeerSetRequest::StemPeer => Ok(self.random_peer_for_stem()),
+        })
+    }
+}
diff --git a/p2p/p2p/src/peer_set/client_wrappers.rs b/p2p/p2p/src/peer_set/client_wrappers.rs
new file mode 100644
index 00000000..97d74931
--- /dev/null
+++ b/p2p/p2p/src/peer_set/client_wrappers.rs
@@ -0,0 +1,86 @@
+use std::{
+    ops::{Deref, DerefMut},
+    sync::{
+        atomic::{AtomicBool, Ordering},
+        Arc,
+    },
+};
+
+use cuprate_p2p_core::{
+    client::{Client, WeakClient},
+    NetworkZone,
+};
+
+/// A client stored in the peer-set.
+pub(super) struct StoredClient<N: NetworkZone> {
+    pub client: Client<N>,
+    /// An [`AtomicBool`] for if the peer is currently downloading blocks.
+    downloading_blocks: Arc<AtomicBool>,
+    /// An [`AtomicBool`] for if the peer is currently being used to stem txs.
+    stem_peer: Arc<AtomicBool>,
+}
+
+impl<N: NetworkZone> StoredClient<N> {
+    pub(super) fn new(client: Client<N>) -> Self {
+        Self {
+            client,
+            downloading_blocks: Arc::new(AtomicBool::new(false)),
+            stem_peer: Arc::new(AtomicBool::new(false)),
+        }
+    }
+
+    /// Returns [`true`] if the [`StoredClient`] is currently downloading blocks.
+    pub(super) fn is_downloading_blocks(&self) -> bool {
+        self.downloading_blocks.load(Ordering::Relaxed)
+    }
+
+    /// Returns [`true`] if the [`StoredClient`] is currently being used to stem txs.
+    pub(super) fn is_a_stem_peer(&self) -> bool {
+        self.stem_peer.load(Ordering::Relaxed)
+    }
+
+    /// Returns a [`ClientDropGuard`] that while it is alive keeps the [`StoredClient`] in the downloading blocks state.
+    pub(super) fn downloading_blocks_guard(&self) -> ClientDropGuard<N> {
+        self.downloading_blocks.store(true, Ordering::Relaxed);
+
+        ClientDropGuard {
+            client: self.client.downgrade(),
+            bool: Arc::clone(&self.downloading_blocks),
+        }
+    }
+
+    /// Returns a [`ClientDropGuard`] that while it is alive keeps the [`StoredClient`] in the stemming peers state.
+    pub(super) fn stem_peer_guard(&self) -> ClientDropGuard<N> {
+        self.stem_peer.store(true, Ordering::Relaxed);
+
+        ClientDropGuard {
+            client: self.client.downgrade(),
+            bool: Arc::clone(&self.stem_peer),
+        }
+    }
+}
+
+/// A [`Drop`] guard for a client returned from the peer-set.
+pub struct ClientDropGuard<N: NetworkZone> {
+    client: WeakClient<N>,
+    bool: Arc<AtomicBool>,
+}
+
+impl<N: NetworkZone> Deref for ClientDropGuard<N> {
+    type Target = WeakClient<N>;
+    fn deref(&self) -> &Self::Target {
+        &self.client
+    }
+}
+
+impl<N: NetworkZone> DerefMut for ClientDropGuard<N> {
+    fn deref_mut(&mut self) -> &mut Self::Target {
+        &mut self.client
+    }
+}
+
+impl<N: NetworkZone> Drop for ClientDropGuard<N> {
+    fn drop(&mut self) {
+        self.bool.store(false, Ordering::Relaxed);
+    }
+}