diff --git a/python/cudaq/_experimental/runtime_endpoint.py b/python/cudaq/_experimental/runtime_endpoint.py
index bde71fcf159..4f2fc6597bf 100644
--- a/python/cudaq/_experimental/runtime_endpoint.py
+++ b/python/cudaq/_experimental/runtime_endpoint.py
@@ -160,9 +160,11 @@ def estimate(self, module: CompiledModule, args: KernelArgs,
                  **kwargs) -> EstimateResult:
         """Estimate the resources a compiled kernel would use.
 
-        Keyword arguments: ``choice``, a callable returning a `bool` that
-        resolves each measurement so that kernels branching on measurement
-        results take a definite path.
+        Keyword arguments include ``choice``, a callable returning a ``bool``
+        that resolves each measurement so that kernels branching on measurement
+        results take a definite path. Additional keyword arguments passed to
+        :func:`cudaq.estimate` are forwarded unchanged for endpoint-specific
+        estimation options.
         """
         ...
 
diff --git a/python/cudaq/runtime/resource_count.py b/python/cudaq/runtime/resource_count.py
index 6bd0638e893..f87f0f2402b 100644
--- a/python/cudaq/runtime/resource_count.py
+++ b/python/cudaq/runtime/resource_count.py
@@ -51,8 +51,11 @@ def estimate(kernel, *args, **kwargs):
         decorator = mk_decorator(kernel)
     processedArgs, module = decorator.prepare_call(*args)
     choice = kwargs.get("choice", None)
+    endpoint_options = {
+        key: value for key, value in kwargs.items() if key != "choice"
+    }
     return cudaq_runtime.estimate_impl(decorator.uniqName, module, choice,
-                                       *processedArgs)
+                                       endpoint_options, *processedArgs)
 
 
 @trace.traced
diff --git a/python/runtime/cudaq/algorithms/py_resource_count.cpp b/python/runtime/cudaq/algorithms/py_resource_count.cpp
index 9234d450c3b..e20d48e7a26 100644
--- a/python/runtime/cudaq/algorithms/py_resource_count.cpp
+++ b/python/runtime/cudaq/algorithms/py_resource_count.cpp
@@ -9,6 +9,7 @@
 #include "py_resource_count.h"
 #include "common/Resources.h"
 #include "common/cudaq_json.h"
+#include "runtime/cudaq/platform/PyRuntimeEndpoint.h"
 #include "runtime/cudaq/platform/py_alt_launch_kernel.h"
 #include "utils/JsonNanobindAdaptors.h"
 #include "utils/OpaqueArguments.h"
@@ -25,7 +26,7 @@ using namespace cudaq;
 static estimate_result
 estimate_impl(const std::string &kernelName, MlirModule kernelMod,
               std::optional<std::function<bool()>> choice,
-              nanobind::args args) {
+              nanobind::dict endpointOptions, nanobind::args args) {
   auto &platform = cudaq::get_platform();
   args = simplifiedValidateInputArguments(args);
 
@@ -47,6 +48,7 @@ estimate_impl(const std::string &kernelName, MlirModule kernelMod,
   estimate_policy policy{
       .kernelName = kernelName,
       .choice = *std::move(choice),
+      .endpointOptions = makePythonEndpointOptions(std::move(endpointOptions)),
   };
   return detail::launch(policy, 0, ctx, platform, [&]() {
     // Pass nullptr for the compiled slot to disable JIT-artifact caching:
@@ -61,8 +63,10 @@ estimate_impl(const std::string &kernelName, MlirModule kernelMod,
 static Resources
 estimate_resources_impl(const std::string &kernelName, MlirModule kernelMod,
                         std::optional<std::function<bool()>> choice,
-                        nanobind::args args) {
-  return estimate_impl(kernelName, kernelMod, choice, args).get_resources();
+                        nanobind::dict endpointOptions, nanobind::args args) {
+  return estimate_impl(kernelName, kernelMod, choice,
+                       std::move(endpointOptions), args)
+      .get_resources();
 }
 
 void cudaq::bindCountResources(nanobind::module_ &mod) {
@@ -117,9 +121,12 @@ void cudaq::bindCountResources(nanobind::module_ &mod) {
 
   mod.def("estimate_impl", estimate_impl, nanobind::arg("kernel_name"),
           nanobind::arg("kernel_mod"), nanobind::arg("choice").none(),
+          nanobind::arg("endpoint_options") = nanobind::dict(),
           nanobind::arg("args"), "See python documentation for estimate.");
   mod.def("estimate_resources_impl", estimate_resources_impl,
           nanobind::arg("kernel_name"), nanobind::arg("kernel_mod"),
-          nanobind::arg("choice").none(), nanobind::arg("args"),
+          nanobind::arg("choice").none(),
+          nanobind::arg("endpoint_options") = nanobind::dict(),
+          nanobind::arg("args"),
           "See python documentation for estimate_resources.");
 }
diff --git a/python/runtime/cudaq/platform/PyRuntimeEndpoint.cpp b/python/runtime/cudaq/platform/PyRuntimeEndpoint.cpp
index a55e4528263..d0232a87ce8 100644
--- a/python/runtime/cudaq/platform/PyRuntimeEndpoint.cpp
+++ b/python/runtime/cudaq/platform/PyRuntimeEndpoint.cpp
@@ -158,6 +158,25 @@ struct PyKernelArgs {
 template <launch_policy Policy>
 struct PyProtocol {};
 
+/// Python-owned options that are opaque to core launch policies.
+struct PythonEndpointOptions final : endpoint_options {
+  explicit PythonEndpointOptions(nanobind::dict options)
+      : options(std::move(options)) {}
+
+  nanobind::dict options;
+};
+
+template <typename Policy>
+static void appendEndpointOptions(const Policy &policy,
+                                  nanobind::dict &kwargs) {
+  auto options =
+      std::dynamic_pointer_cast<PythonEndpointOptions>(policy.endpointOptions);
+  if (!options)
+    return;
+  for (auto [key, value] : options->options)
+    kwargs[key] = value;
+}
+
 template <>
 struct PyProtocol<sample_policy> {
   static constexpr const char *Method = "sample";
@@ -261,6 +280,7 @@ pyLaunch(std::any &impl, const Policy &policy, const CompiledModule &module,
                                nanobind::rv_policy::move);
 
   auto kwargs = Protocol::kwargs(policy);
+  appendEndpointOptions(policy, kwargs);
   auto result = obj.attr(Protocol::Method)(pyModule, pyArgs, **kwargs);
 
   if (!nanobind::isinstance<Result>(result)) {
@@ -278,6 +298,25 @@ pyLaunch(std::any &impl, const Policy &policy, const CompiledModule &module,
   return nanobind::cast<typename Policy::result_type>(result);
 }
 
+std::shared_ptr<endpoint_options>
+cudaq::makePythonEndpointOptions(nanobind::dict options) {
+  if (options.empty())
+    return {};
+
+  auto optionsDestructor = +[](endpoint_options *base) {
+    auto *options = static_cast<PythonEndpointOptions *>(base);
+    if (!Py_IsInitialized()) {
+      (void)options->options.release();
+      delete options;
+      return;
+    }
+    nanobind::gil_scoped_acquire gil;
+    delete options;
+  };
+  return std::shared_ptr<endpoint_options>(
+      new PythonEndpointOptions(std::move(options)), optionsDestructor);
+}
+
 static bool getAttrOrDefault(const nanobind::object &obj, const char *attr,
                              bool defaultValue) {
   if (nanobind::hasattr(obj, attr))
diff --git a/python/runtime/cudaq/platform/PyRuntimeEndpoint.h b/python/runtime/cudaq/platform/PyRuntimeEndpoint.h
index e4a0a024edc..43ae5e7ba9c 100644
--- a/python/runtime/cudaq/platform/PyRuntimeEndpoint.h
+++ b/python/runtime/cudaq/platform/PyRuntimeEndpoint.h
@@ -9,6 +9,7 @@
 #pragma once
 
 #include "cudaq/Target/RuntimeEndpoint.h"
+#include <memory>
 #include <nanobind/nanobind.h>
 
 namespace cudaq {
@@ -16,4 +17,8 @@ namespace cudaq {
 /// Create python bindings for C++ code in this compilation unit.
 void bindRuntimeEndpoint(nanobind::module_ &mod);
 
+/// Preserve Python-only endpoint keyword arguments on a launch policy.
+std::shared_ptr<endpoint_options>
+makePythonEndpointOptions(nanobind::dict options);
+
 } // namespace cudaq
diff --git a/python/tests/backends/test_experimental_runtime_endpoint.py b/python/tests/backends/test_experimental_runtime_endpoint.py
index 5ee249e93ad..19e91738d11 100644
--- a/python/tests/backends/test_experimental_runtime_endpoint.py
+++ b/python/tests/backends/test_experimental_runtime_endpoint.py
@@ -220,6 +220,18 @@ def test_estimate_forwards_the_choice_function():
     assert kwargs["choice"]() is True
 
 
+def test_estimate_forwards_endpoint_options():
+    endpoint = DemoEndpoint()
+    set_runtime_endpoint(endpoint)
+
+    marker = object()
+    cudaq.estimate(kernel, 1, [1, 2, 3], tier="logical", marker=marker)
+
+    _, _, kwargs = endpoint.calls[0]
+    assert kwargs["tier"] == "logical"
+    assert kwargs["marker"] is marker
+
+
 def test_estimate_resources_launch():
     endpoint = DemoEndpoint()
     set_runtime_endpoint(endpoint)
diff --git a/runtime/cudaq/algorithms/dem/policy.h b/runtime/cudaq/algorithms/dem/policy.h
index c3150d8630f..1684b4acc6d 100644
--- a/runtime/cudaq/algorithms/dem/policy.h
+++ b/runtime/cudaq/algorithms/dem/policy.h
@@ -11,6 +11,7 @@
 #include "common/CompileOptions.h"
 #include "cudaq/algorithms/dem/options.h"
 #include "cudaq/algorithms/dem/result.h"
+#include "cudaq/algorithms/endpoint_options.h"
 #include <string>
 
 namespace cudaq {
@@ -24,6 +25,7 @@ struct dem_policy {
   dem_options options;
   std::string kernelName;
   const noise_model *noiseModel = nullptr;
+  std::shared_ptr<endpoint_options> endpointOptions;
 
   friend CompileOptions get_compile_options_impl(const dem_policy &);
 };
diff --git a/runtime/cudaq/algorithms/endpoint_options.h b/runtime/cudaq/algorithms/endpoint_options.h
new file mode 100644
index 00000000000..0ff0b1d8d93
--- /dev/null
+++ b/runtime/cudaq/algorithms/endpoint_options.h
@@ -0,0 +1,21 @@
+/****************************************************************-*- C++ -*-****
+ * Copyright (c) 2026 NVIDIA Corporation & Affiliates.                         *
+ * All rights reserved.                                                        *
+ *                                                                             *
+ * This source code and the accompanying materials are made available under    *
+ * the terms of the Apache License 2.0 which accompanies this distribution.    *
+ ******************************************************************************/
+
+#pragma once
+
+#include <memory>
+
+namespace cudaq {
+
+/// Type-erased, endpoint-specific options attached to a launch policy.
+/// Core policies remain independent of the language used by an endpoint.
+struct endpoint_options {
+  virtual ~endpoint_options() = default;
+};
+
+} // namespace cudaq
diff --git a/runtime/cudaq/algorithms/estimate/policy.h b/runtime/cudaq/algorithms/estimate/policy.h
index a9b49ba1456..f130d9a5383 100644
--- a/runtime/cudaq/algorithms/estimate/policy.h
+++ b/runtime/cudaq/algorithms/estimate/policy.h
@@ -9,6 +9,7 @@
 #pragma once
 
 #include "common/CompileOptions.h"
+#include "cudaq/algorithms/endpoint_options.h"
 #include "cudaq/algorithms/estimate/result.h"
 #include <functional>
 #include <string>
@@ -28,6 +29,9 @@ struct estimate_policy {
   /// follow when the kernel branches on a measurement result.
   std::function<bool()> choice;
 
+  /// Options for a runtime endpoint that CUDA-Q itself does not interpret.
+  std::shared_ptr<endpoint_options> endpointOptions;
+
   friend CompileOptions get_compile_options_impl(const estimate_policy &);
 };
 
diff --git a/runtime/cudaq/algorithms/observe/policy.h b/runtime/cudaq/algorithms/observe/policy.h
index 0a593e3588d..a499c7f39e5 100644
--- a/runtime/cudaq/algorithms/observe/policy.h
+++ b/runtime/cudaq/algorithms/observe/policy.h
@@ -11,6 +11,7 @@
 #include "common/CompileOptions.h"
 #include "common/Future.h"
 #include "common/ObserveResult.h"
+#include "cudaq/algorithms/endpoint_options.h"
 #include "cudaq/algorithms/observe/options.h"
 #include "cudaq/operators.h"
 
@@ -44,6 +45,8 @@ struct observe_policy {
 
   mutable bool canHandleObserve = false;
 
+  std::shared_ptr<endpoint_options> endpointOptions;
+
   friend observe_result
   finalize_execution_manager_impl(ExecutionManager &mgr,
                                   const observe_policy &policy,
diff --git a/runtime/cudaq/algorithms/sample/policy.h b/runtime/cudaq/algorithms/sample/policy.h
index 83b01f2d608..939c7e687af 100644
--- a/runtime/cudaq/algorithms/sample/policy.h
+++ b/runtime/cudaq/algorithms/sample/policy.h
@@ -11,6 +11,7 @@
 #include "common/CompileOptions.h"
 #include "common/Future.h"
 #include "common/SampleResult.h"
+#include "cudaq/algorithms/endpoint_options.h"
 #include "cudaq/algorithms/sample/options.h"
 
 namespace nvqir {
@@ -42,6 +43,8 @@ struct sample_policy {
 
   mutable const noise_model *noiseModel = nullptr;
 
+  std::shared_ptr<endpoint_options> endpointOptions;
+
   friend sample_result
   finalize_execution_manager_impl(ExecutionManager &mgr,
                                   const sample_policy &policy);