diff --git a/Cargo.lock b/Cargo.lock
index e35a6cd37a..ad90213848 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -342,6 +342,7 @@ dependencies = [
  "azure_core_test",
  "azure_identity",
  "azure_security_keyvault_test",
+ "clap",
  "criterion",
  "futures",
  "rand",
diff --git a/sdk/core/azure_core_test/src/lib.rs b/sdk/core/azure_core_test/src/lib.rs
index c6cde267e9..ff9d399cd7 100644
--- a/sdk/core/azure_core_test/src/lib.rs
+++ b/sdk/core/azure_core_test/src/lib.rs
@@ -33,7 +33,8 @@ pub struct TestContext {
 }
 
 impl TestContext {
-    pub(crate) fn new(
+    #[doc(hidden)]
+    pub fn new(
         crate_dir: &'static str,
         module_dir: &'static str,
         name: &'static str,
diff --git a/sdk/keyvault/azure_security_keyvault_keys/Cargo.toml b/sdk/keyvault/azure_security_keyvault_keys/Cargo.toml
index c1b5ad64fc..1c32fc9cee 100644
--- a/sdk/keyvault/azure_security_keyvault_keys/Cargo.toml
+++ b/sdk/keyvault/azure_security_keyvault_keys/Cargo.toml
@@ -28,6 +28,7 @@ azure_core_test = { workspace = true, features = [
 ] }
 azure_identity.workspace = true
 azure_security_keyvault_test = { path = "../azure_security_keyvault_test" }
+clap.workspace = true
 criterion.workspace = true
 rand.workspace = true
 sha2.workspace = true
@@ -42,3 +43,7 @@ workspace = true
 [[bench]]
 name = "benchmarks"
 harness = false
+
+[[bench]]
+name = "perf"
+harness = false
diff --git a/sdk/keyvault/azure_security_keyvault_keys/benches/perf.rs b/sdk/keyvault/azure_security_keyvault_keys/benches/perf.rs
new file mode 100644
index 0000000000..acdeb20570
--- /dev/null
+++ b/sdk/keyvault/azure_security_keyvault_keys/benches/perf.rs
@@ -0,0 +1,109 @@
+// Copyright (c) Microsoft Corporation. All rights reserved.
+// Licensed under the MIT License.
+
+use azure_core::{
+    error::{Error, ErrorKind, Result},
+    http::RequestContent,
+};
+use azure_core_test::TestContext;
+use azure_security_keyvault_keys::{
+    models::{CreateKeyParameters, CurveName, KeyType},
+    KeyClient,
+};
+use clap::Parser;
+use futures::TryStreamExt as _;
+use std::sync::Arc;
+
+// Could also consider a PerfContext, but we'd still want to probably forward methods from TestContext so may not be worth it
+// given this is purely an internal API: no reason to complicate design and make maintenance more expensive.
+async fn list_keys(ctx: TestContext, args: &Args) -> Result<()> {
+    let recording = ctx.recording();
+    let endpoint = recording
+        .var_opt("AZURE_KEYVAULT_URL", None)
+        .ok_or_else(|| Error::new(ErrorKind::Other, "expected AZURE_KEYVAULT_URL"))?;
+
+    let client = Arc::new(KeyClient::new(&endpoint, recording.credential(), None)?);
+
+    // Setup
+    let params = CreateKeyParameters {
+        kty: Some(KeyType::EC),
+        curve: Some(CurveName::P256),
+        ..Default::default()
+    };
+    let params: RequestContent<CreateKeyParameters> = params.try_into()?;
+    for i in 0..args.count {
+        client
+            .create_key(&format!("key{i}"), params.clone(), None)
+            .await?;
+    }
+
+    // Run
+    // This is a place where we could also define some helpers that takes a capture.
+    for _ in 0..args.common.parallel {
+        let client = client.clone();
+        tokio::spawn(async move {
+            let mut count = 0;
+            let mut pager = client.list_key_properties(None).expect("expected pager");
+            while let Ok(Some(_)) = pager.try_next().await {
+                count += 1;
+            }
+            println!("listed {count} keys");
+        });
+    }
+
+    Ok(())
+}
+
+// We'd have a functional macro or something to define this, just like criterion.
+// The advantage here is that we set `harness = false` but it's not criterion's `main` so it ignores it.
+// Would run with:
+//
+//     cargo bench -p azure_security_keyvault_keys --bench perf -- --parallel 64 --count 50
+#[tokio::main]
+async fn main() {
+    let args = Args::parse();
+
+    // Like criterion, we probably should run each async test serially.
+    // That said, it might also be that the perf framework calls each method individually, in which case we can use a separate file for each test
+    // and you don't need a separate function.
+    {
+        // We'd actually want to create a new TestContext (or PerfContext) in azure_core_test and have it do the work
+        // must like it does for tests. In this case, we don't want it to bootstrap the TestProxy, though, but still find the .env file
+        // and all that. So it probably needs to be a separate proc macro or something. If we assume a single perf test per module (file),
+        // it could generate the `main` and everything. Imagine a declaration like:
+        //
+        // #[perf::test]
+        // async fn list_secrets(ctx: PerfContext<Args>) -> azure_core::Result<()> {
+        //     todo!()
+        // }
+        // #[derive(clap::Parser)]
+        // ...
+        //
+        // And it does all of this, basically. `PerfContext<T>` could constrain `T` to a `PerfArgs` trait or something
+        // so if no specific args are needed, we can default it to a `CommonArgs` as shown below, where `PerfArgs: Parser`, which `CommonArgs` implements.
+        let ctx =
+            TestContext::new(env!("CARGO_MANIFEST_DIR"), module_path!(), "list_keys").unwrap();
+        list_keys(ctx, &args).await.unwrap();
+    }
+}
+
+// Would be defined in azure_core_test or something common.
+#[derive(Debug, Default, clap::Args)]
+struct CommonArgs {
+    #[arg(long)]
+    bench: bool,
+
+    #[arg(long, default_value_t = 1)]
+    parallel: usize,
+}
+
+// Defined by each perf test module.
+#[derive(Debug, Default, Parser)]
+#[command(about, long_about = None, version)]
+struct Args {
+    #[command(flatten)]
+    common: CommonArgs,
+
+    #[arg(long, default_value_t = 10)]
+    count: usize,
+}