diff --git a/buildlib/azure-pipelines-pr.yml b/buildlib/azure-pipelines-pr.yml
index d446fe72ea7..44ccf60fa5f 100644
--- a/buildlib/azure-pipelines-pr.yml
+++ b/buildlib/azure-pipelines-pr.yml
@@ -31,5 +31,13 @@ pr:
     - buildlib/tools/perf_results.py
     - buildlib/tools/perf-common.yml
 
+resources:
+  repositories:
+    - repository: ucxx
+      type: github
+      name: rapidsai/ucxx
+      endpoint: Mellanox-lab
+      ref: refs/heads/main
+
 extends:
   template: pr/main.yml
diff --git a/buildlib/dockers/rapidsai-ci-conda.Dockerfile b/buildlib/dockers/rapidsai-ci-conda.Dockerfile
new file mode 100644
index 00000000000..3e114744552
--- /dev/null
+++ b/buildlib/dockers/rapidsai-ci-conda.Dockerfile
@@ -0,0 +1,7 @@
+# Azure Pipelines wrapper around rapidsai/ci-conda.
+# Opens /opt/conda so the Azure-injected step user can use conda/python.
+
+ARG BASE_IMAGE=rapidsai/ci-conda:26.06-latest
+FROM ${BASE_IMAGE}
+
+RUN chmod -R o+rwX /opt/conda
diff --git a/buildlib/dockers/rapidsai-ci-wheel.Dockerfile b/buildlib/dockers/rapidsai-ci-wheel.Dockerfile
new file mode 100644
index 00000000000..52a82dada8a
--- /dev/null
+++ b/buildlib/dockers/rapidsai-ci-wheel.Dockerfile
@@ -0,0 +1,7 @@
+# Azure Pipelines wrapper around rapidsai/ci-wheel.
+# Opens /pyenv so the Azure-injected step user can write shims.
+
+ARG BASE_IMAGE=rapidsai/ci-wheel:26.06-latest
+FROM ${BASE_IMAGE}
+
+RUN chmod -R o+rwX /pyenv
diff --git a/buildlib/pr/main.yml b/buildlib/pr/main.yml
index a47f61052b9..f7b06fdd600 100644
--- a/buildlib/pr/main.yml
+++ b/buildlib/pr/main.yml
@@ -249,6 +249,20 @@ resources:
     - container: centos10stream
       image: rdmz-harbor.rdmz.labs.mlnx/hpcx/x86_64/centos10stream/builder:inbox
       options: $(DOCKER_OPT_ARGS) $(DOCKER_OPT_VOLUMES)
+    - container: ucxx_rapidsai_ci_conda
+      # Thin wrapper of rapidsai/ci-conda; see buildlib/dockers/rapidsai-ci-conda.Dockerfile.
+      image: rdmz-harbor.rdmz.labs.mlnx/ucx/rapidsai-ci-conda:26.06-azp-1
+      options: $(DOCKER_OPT_ARGS) $(DOCKER_OPT_VOLUMES)
+    - container: ucxx_rapidsai_ci_conda_gpu
+      # Same image as above, +GPU. Bind-mount host MPS dirs so the container's
+      # CUDA client can reach the host MPS daemon (host GPUs are in
+      # Exclusive_Process compute mode and need MPS for shared access).
+      image: rdmz-harbor.rdmz.labs.mlnx/ucx/rapidsai-ci-conda:26.06-azp-1
+      options: $(DOCKER_OPT_ARGS) $(DOCKER_OPT_VOLUMES) $(DOCKER_OPT_GPU) -v /tmp/nvidia-mps:/tmp/nvidia-mps
+    - container: ucxx_rapidsai_ci_wheel
+      # Thin wrapper of rapidsai/ci-wheel; see buildlib/dockers/rapidsai-ci-wheel.Dockerfile.
+      image: rdmz-harbor.rdmz.labs.mlnx/ucx/rapidsai-ci-wheel:26.06-azp-1
+      options: $(DOCKER_OPT_ARGS) $(DOCKER_OPT_VOLUMES)
 
 stages:
   - stage: Codestyle
@@ -256,16 +270,19 @@ stages:
     - template: codestyle.yml
 
   - stage: Basic_compile
+    condition: false  # temp: skip non-UCXX stage; restore before merge
     dependsOn: []
     jobs:
     - template: basic_compile.yml
 
   - stage: Static_check
+    condition: false  # temp: skip non-UCXX stage; restore before merge
     dependsOn: [Basic_compile]
     jobs:
     - template: static_checks.yml
 
   - stage: Build
+    condition: false  # temp: skip non-UCXX stage; restore before merge
     dependsOn: [Basic_compile]
     jobs:
       - template: build_job.yml
@@ -283,6 +300,7 @@ stages:
           displayName: Build on aarch64
 
   - stage: ucx_perftest_mad_rte
+    condition: false  # temp: skip non-UCXX stage; restore before merge
     dependsOn: [Basic_compile]
     displayName: ucx_perftest over MAD RTE
     lockBehavior: sequential
@@ -292,6 +310,7 @@ stages:
     - template: mad_tests.yml
 
   - stage: WireCompat
+    condition: false  # temp: skip non-UCXX stage; restore before merge
     dependsOn: [Basic_compile]
     jobs:
     - template: wire_compat.yml
@@ -323,6 +342,7 @@ stages:
     #    demands: ucx_iodemo -equals yes
 
   - stage: Coverity
+    condition: false  # temp: skip non-UCXX stage; restore before merge
     dependsOn: [Static_check]
     jobs:
       - template: coverity.yml
@@ -330,7 +350,118 @@ stages:
           demands: ucx_docker -equals yes
           container: coverity_rh7
 
+  - template: ucxx_tests_stage.yml
+    parameters:
+      dependsOn: []  # temp: skip Static_check to test stage faster on this PR; restore [Static_check] before merge
+
+  - stage: UCXX_build
+    dependsOn: []  # temp: skip Static_check while iterating Phase 2; restore [Static_check] before merge
+    jobs:
+      # 2.1: conda-cpp-build. Mirrors upstream conda-cpp-build matrix (4 legs).
+      - template: ucxx_conda_cpp_build.yml
+        parameters:
+          container: ucxx_rapidsai_ci_conda
+          slices:
+            - name: x86_64_cuda12_py311
+              demands: ucx_docker -equals yes
+              rapids_cuda_version: '12.9.1'
+              rapids_py_version: '3.11'
+            - name: x86_64_cuda13_py311
+              demands: ucx_docker -equals yes
+              rapids_cuda_version: '13.2.0'
+              rapids_py_version: '3.11'
+            - name: aarch64_cuda12_py311
+              demands: ucx_arm64
+              rapids_cuda_version: '12.9.1'
+              rapids_py_version: '3.11'
+            - name: aarch64_cuda13_py311
+              demands: ucx_arm64
+              rapids_cuda_version: '13.2.0'
+              rapids_py_version: '3.11'
+      # 2.2: conda-python-build. Depends on 2.1's artifacts (per-slice).
+      - template: ucxx_conda_python_build.yml
+        parameters:
+          container: ucxx_rapidsai_ci_conda
+          slices:
+            - name: x86_64_cuda12_py311
+              demands: ucx_docker -equals yes
+              rapids_cuda_version: '12.9.1'
+              rapids_py_version: '3.11'
+            - name: x86_64_cuda13_py311
+              demands: ucx_docker -equals yes
+              rapids_cuda_version: '13.2.0'
+              rapids_py_version: '3.11'
+            - name: aarch64_cuda12_py311
+              demands: ucx_arm64
+              rapids_cuda_version: '12.9.1'
+              rapids_py_version: '3.11'
+            - name: aarch64_cuda13_py311
+              demands: ucx_arm64
+              rapids_cuda_version: '13.2.0'
+              rapids_py_version: '3.11'
+      # 2.3: wheel-build-libucxx. Upstream filter = max_by(PY_VER) per (ARCH, CUDA major).
+      - template: ucxx_wheel_libucxx_build.yml
+        parameters:
+          container: ucxx_rapidsai_ci_wheel
+          slices:
+            - name: x86_64_cuda12_py314
+              demands: ucx_docker -equals yes
+              rapids_cuda_version: '12.9.1'
+              rapids_py_version: '3.14'
+            - name: x86_64_cuda13_py314
+              demands: ucx_docker -equals yes
+              rapids_cuda_version: '13.0.2'
+              rapids_py_version: '3.14'
+            - name: aarch64_cuda12_py314
+              demands: ucx_arm64
+              rapids_cuda_version: '12.9.1'
+              rapids_py_version: '3.14'
+            - name: aarch64_cuda13_py314
+              demands: ucx_arm64
+              rapids_cuda_version: '13.0.2'
+              rapids_py_version: '3.14'
+      # 2.4: wheel-build-ucxx. Upstream filter = min_by(PY_VER) per (CUDA, ARCH).
+      # Each ucxx wheel slice consumes a libucxx wheel slice; pair by (ARCH, CUDA).
+      - template: ucxx_wheel_ucxx_build.yml
+        parameters:
+          container: ucxx_rapidsai_ci_wheel
+          slices:
+            - name: x86_64_cuda12_py311
+              libucxx_slice: x86_64_cuda12_py314
+              demands: ucx_docker -equals yes
+              rapids_cuda_version: '12.9.1'
+              rapids_py_version: '3.11'
+            - name: x86_64_cuda13_py311
+              libucxx_slice: x86_64_cuda13_py314
+              demands: ucx_docker -equals yes
+              rapids_cuda_version: '13.0.2'
+              rapids_py_version: '3.11'
+            - name: aarch64_cuda12_py311
+              libucxx_slice: aarch64_cuda12_py314
+              demands: ucx_arm64
+              rapids_cuda_version: '12.9.1'
+              rapids_py_version: '3.11'
+            - name: aarch64_cuda13_py311
+              libucxx_slice: aarch64_cuda13_py314
+              demands: ucx_arm64
+              rapids_cuda_version: '13.0.2'
+              rapids_py_version: '3.11'
+      # 2.5: wheel-build-distributed-ucxx. Pure-python; upstream emits 2 amd64 legs.
+      - template: ucxx_wheel_distributed_ucxx_build.yml
+        parameters:
+          container: ucxx_rapidsai_ci_wheel
+          slices:
+            - name: x86_64_cuda12_py314
+              demands: ucx_docker -equals yes
+              rapids_cuda_version: '12.9.1'
+              rapids_py_version: '3.14'
+            - name: x86_64_cuda13_py314
+              demands: ucx_docker -equals yes
+              rapids_cuda_version: '13.0.2'
+              rapids_py_version: '3.14'
+
   - stage: Tests
+    condition: false  # temp: skip non-UCXX stage; restore before merge
     dependsOn: [Basic_compile]
     jobs:
     - template: tests.yml
@@ -410,6 +541,7 @@ stages:
             demands: ucx_gpu -equals yes
 
   - stage: Build_Static
+    condition: false  # temp: skip non-UCXX stage; restore before merge
     dependsOn: [Basic_compile]
     jobs:
       - job: build_source
@@ -456,12 +588,14 @@ stages:
 
 
   - stage: Cuda
+    condition: false  # temp: skip non-UCXX stage; restore before merge
     dependsOn: [Basic_compile]
     jobs:
       - template: cuda/cuda.yml
 
 
   - stage: AddressSanitizer
+    condition: false  # temp: skip non-UCXX stage; restore before merge
     dependsOn: [Static_check]
     jobs:
     - template: tests.yml
diff --git a/buildlib/pr/ucxx_conda_cpp_build.yml b/buildlib/pr/ucxx_conda_cpp_build.yml
new file mode 100644
index 00000000000..03674583bc2
--- /dev/null
+++ b/buildlib/pr/ucxx_conda_cpp_build.yml
@@ -0,0 +1,40 @@
+parameters:
+  container: ucxx_rapidsai_ci_conda
+  slices: []
+
+jobs:
+  - ${{ each slice in parameters.slices }}:
+    - job: ucxx_conda_cpp_build_${{ slice.name }}
+      workspace:
+        clean: all
+      pool:
+        name: MLNX
+        demands: ${{ slice.demands }}
+      displayName: 'UCXX conda-cpp-build (${{ slice.name }})'
+      container: ${{ parameters.container }}
+      timeoutInMinutes: 60
+      variables:
+        UCX_DIR: $(Agent.BuildDirectory)/ucx
+        UCXX_DIR: $(Agent.BuildDirectory)/ucxx
+        RAPIDS_CUDA_VERSION: ${{ slice.rapids_cuda_version }}
+        RAPIDS_PY_VERSION: ${{ slice.rapids_py_version }}
+        RAPIDS_BLD_OUTPUT_DIR: $(Build.ArtifactStagingDirectory)
+
+      steps:
+        - checkout: self
+          path: ucx
+          fetchDepth: 100
+          retryCountOnTaskFailure: 5
+
+        - checkout: ucxx
+          path: ucxx
+          retryCountOnTaskFailure: 5
+
+        - bash: bash $(UCX_DIR)/buildlib/tools/build_ucxx.sh conda_cpp
+          displayName: Build UCXX conda C++ package
+
+        - task: PublishBuildArtifacts@1
+          displayName: Publish conda-cpp artifact
+          inputs:
+            pathToPublish: $(RAPIDS_BLD_OUTPUT_DIR)
+            artifactName: ucxx-conda-cpp-${{ slice.name }}
diff --git a/buildlib/pr/ucxx_conda_python_build.yml b/buildlib/pr/ucxx_conda_python_build.yml
new file mode 100644
index 00000000000..da5565be1e3
--- /dev/null
+++ b/buildlib/pr/ucxx_conda_python_build.yml
@@ -0,0 +1,53 @@
+parameters:
+  container: ucxx_rapidsai_ci_conda
+  slices: []
+
+jobs:
+  - ${{ each slice in parameters.slices }}:
+    - job: ucxx_conda_python_build_${{ slice.name }}
+      workspace:
+        clean: all
+      pool:
+        name: MLNX
+        demands: ${{ slice.demands }}
+      displayName: 'UCXX conda-python-build (${{ slice.name }})'
+      container: ${{ parameters.container }}
+      timeoutInMinutes: 60
+      dependsOn: ucxx_conda_cpp_build_${{ slice.name }}
+      variables:
+        UCX_DIR: $(Agent.BuildDirectory)/ucx
+        UCXX_DIR: $(Agent.BuildDirectory)/ucxx
+        RAPIDS_CUDA_VERSION: ${{ slice.rapids_cuda_version }}
+        RAPIDS_PY_VERSION: ${{ slice.rapids_py_version }}
+        RAPIDS_BLD_OUTPUT_DIR: $(Build.ArtifactStagingDirectory)
+
+      steps:
+        - checkout: self
+          path: ucx
+          fetchDepth: 100
+          retryCountOnTaskFailure: 5
+
+        - checkout: ucxx
+          path: ucxx
+          retryCountOnTaskFailure: 5
+
+        - task: DownloadBuildArtifacts@1
+          displayName: Fetch conda-cpp artifact
+          inputs:
+            buildType: current
+            artifactName: ucxx-conda-cpp-${{ slice.name }}
+            downloadPath: $(System.DefaultWorkingDirectory)/_dl
+
+        - bash: |
+            rm -rf "$(RAPIDS_BLD_OUTPUT_DIR)"
+            mv "$(System.DefaultWorkingDirectory)/_dl/ucxx-conda-cpp-${{ slice.name }}" "$(RAPIDS_BLD_OUTPUT_DIR)"
+          displayName: Stage conda-cpp artifact
+
+        - bash: bash $(UCX_DIR)/buildlib/tools/build_ucxx.sh conda_python
+          displayName: Build UCXX conda Python package
+
+        - task: PublishBuildArtifacts@1
+          displayName: Publish conda-python artifact
+          inputs:
+            pathToPublish: $(RAPIDS_BLD_OUTPUT_DIR)
+            artifactName: ucxx-conda-python-${{ slice.name }}
diff --git a/buildlib/pr/ucxx_tests_stage.yml b/buildlib/pr/ucxx_tests_stage.yml
new file mode 100644
index 00000000000..850f5c53015
--- /dev/null
+++ b/buildlib/pr/ucxx_tests_stage.yml
@@ -0,0 +1,85 @@
+parameters:
+  dependsOn: [Static_check]
+  cpu_container: ucxx_rapidsai_ci_conda
+  gpu_container: ucxx_rapidsai_ci_conda_gpu
+  slices:
+    # CPU coverage: cuda x arch x py grid; builds conda C++/Python pkgs and
+    # runs UCXX gtests on host transports (CUDA-free subset).
+    - name: x86_64_cuda12_py311
+      gpu: false
+      demands: ucx_docker -equals yes
+      rapids_cuda_version: '12.9.1'
+      rapids_py_version: '3.11'
+    - name: x86_64_cuda13_py311
+      gpu: false
+      demands: ucx_docker -equals yes
+      rapids_cuda_version: '13.2.0'
+      rapids_py_version: '3.11'
+    - name: aarch64_cuda12_py311
+      gpu: false
+      demands: ucx_arm64
+      rapids_cuda_version: '12.9.1'
+      rapids_py_version: '3.11'
+    - name: aarch64_cuda13_py311
+      gpu: false
+      demands: ucx_arm64
+      rapids_cuda_version: '13.2.0'
+      rapids_py_version: '3.11'
+    # GPU coverage: x86_64 only (MLNX ucx_gpu_test pool has no arm64 hosts);
+    # py 3.12 + 3.13 to exercise both supported toolchains.
+    - name: amd64_cuda1302_py312
+      gpu: true
+      demands: ucx_gpu_test -equals yes
+      rapids_cuda_version: '13.0.2'
+      rapids_py_version: '3.12'
+    - name: amd64_cuda1322_py313
+      gpu: true
+      demands: ucx_gpu_test -equals yes
+      rapids_cuda_version: '13.2.0'
+      rapids_py_version: '3.13'
+
+stages:
+  - stage: UCXX_tests
+    dependsOn: ${{ parameters.dependsOn }}
+    jobs:
+      - ${{ each slice in parameters.slices }}:
+        - job: ucxx_tests_${{ slice.name }}
+          workspace:
+            clean: all
+          pool:
+            name: MLNX
+            demands: ${{ slice.demands }}
+          ${{ if eq(slice.gpu, true) }}:
+            displayName: 'UCXX GPU tests (${{ slice.name }})'
+            container: ${{ parameters.gpu_container }}
+            timeoutInMinutes: 120
+          ${{ if eq(slice.gpu, false) }}:
+            displayName: 'UCXX tests (${{ slice.name }})'
+            container: ${{ parameters.cpu_container }}
+            timeoutInMinutes: 90
+          variables:
+            IS_GPU: ${{ slice.gpu }}
+            RAPIDS_CUDA_VERSION: ${{ slice.rapids_cuda_version }}
+            RAPIDS_PY_VERSION: ${{ slice.rapids_py_version }}
+            UCX_DIR: $(Agent.BuildDirectory)/ucx
+            UCXX_DIR: $(Agent.BuildDirectory)/ucxx
+
+          steps:
+            - checkout: self
+              path: ucx
+              fetchDepth: 100
+              retryCountOnTaskFailure: 5
+
+            - checkout: ucxx
+              path: ucxx
+              retryCountOnTaskFailure: 5
+
+            - bash: bash $(UCX_DIR)/buildlib/tools/test_ucxx.sh build
+              displayName: Build UCXX
+
+            - bash: bash $(UCX_DIR)/buildlib/tools/test_ucxx.sh test_cpp
+              displayName: Run UCXX C++ tests
+
+            - ${{ if eq(slice.gpu, true) }}:
+              - bash: bash $(UCX_DIR)/buildlib/tools/test_ucxx.sh test_python
+                displayName: Run UCXX Python tests
diff --git a/buildlib/pr/ucxx_wheel_distributed_ucxx_build.yml b/buildlib/pr/ucxx_wheel_distributed_ucxx_build.yml
new file mode 100644
index 00000000000..cbbafceef12
--- /dev/null
+++ b/buildlib/pr/ucxx_wheel_distributed_ucxx_build.yml
@@ -0,0 +1,40 @@
+parameters:
+  container: ucxx_rapidsai_ci_wheel
+  slices: []
+
+jobs:
+  - ${{ each slice in parameters.slices }}:
+    - job: ucxx_wheel_dist_ucxx_build_${{ slice.name }}
+      workspace:
+        clean: all
+      pool:
+        name: MLNX
+        demands: ${{ slice.demands }}
+      displayName: 'UCXX wheel-build-distributed-ucxx (${{ slice.name }})'
+      container: ${{ parameters.container }}
+      timeoutInMinutes: 45
+      variables:
+        UCX_DIR: $(Agent.BuildDirectory)/ucx
+        UCXX_DIR: $(Agent.BuildDirectory)/ucxx
+        RAPIDS_CUDA_VERSION: ${{ slice.rapids_cuda_version }}
+        RAPIDS_PY_VERSION: ${{ slice.rapids_py_version }}
+        RAPIDS_BLD_OUTPUT_DIR: $(Build.ArtifactStagingDirectory)
+
+      steps:
+        - checkout: self
+          path: ucx
+          fetchDepth: 100
+          retryCountOnTaskFailure: 5
+
+        - checkout: ucxx
+          path: ucxx
+          retryCountOnTaskFailure: 5
+
+        - bash: bash $(UCX_DIR)/buildlib/tools/build_ucxx.sh wheel_distributed_ucxx
+          displayName: Build distributed-ucxx wheel
+
+        - task: PublishBuildArtifacts@1
+          displayName: Publish distributed-ucxx wheel artifact
+          inputs:
+            pathToPublish: $(RAPIDS_BLD_OUTPUT_DIR)
+            artifactName: ucxx-wheel-distributed-ucxx-${{ slice.name }}
diff --git a/buildlib/pr/ucxx_wheel_libucxx_build.yml b/buildlib/pr/ucxx_wheel_libucxx_build.yml
new file mode 100644
index 00000000000..28525d844aa
--- /dev/null
+++ b/buildlib/pr/ucxx_wheel_libucxx_build.yml
@@ -0,0 +1,40 @@
+parameters:
+  container: ucxx_rapidsai_ci_wheel
+  slices: []
+
+jobs:
+  - ${{ each slice in parameters.slices }}:
+    - job: ucxx_wheel_libucxx_build_${{ slice.name }}
+      workspace:
+        clean: all
+      pool:
+        name: MLNX
+        demands: ${{ slice.demands }}
+      displayName: 'UCXX wheel-build-libucxx (${{ slice.name }})'
+      container: ${{ parameters.container }}
+      timeoutInMinutes: 60
+      variables:
+        UCX_DIR: $(Agent.BuildDirectory)/ucx
+        UCXX_DIR: $(Agent.BuildDirectory)/ucxx
+        RAPIDS_CUDA_VERSION: ${{ slice.rapids_cuda_version }}
+        RAPIDS_PY_VERSION: ${{ slice.rapids_py_version }}
+        RAPIDS_BLD_OUTPUT_DIR: $(Build.ArtifactStagingDirectory)
+
+      steps:
+        - checkout: self
+          path: ucx
+          fetchDepth: 100
+          retryCountOnTaskFailure: 5
+
+        - checkout: ucxx
+          path: ucxx
+          retryCountOnTaskFailure: 5
+
+        - bash: bash $(UCX_DIR)/buildlib/tools/build_ucxx.sh wheel_libucxx
+          displayName: Build libucxx wheel
+
+        - task: PublishBuildArtifacts@1
+          displayName: Publish libucxx wheel artifact
+          inputs:
+            pathToPublish: $(RAPIDS_BLD_OUTPUT_DIR)
+            artifactName: ucxx-wheel-libucxx-${{ slice.name }}
diff --git a/buildlib/pr/ucxx_wheel_ucxx_build.yml b/buildlib/pr/ucxx_wheel_ucxx_build.yml
new file mode 100644
index 00000000000..35d0f3150ac
--- /dev/null
+++ b/buildlib/pr/ucxx_wheel_ucxx_build.yml
@@ -0,0 +1,54 @@
+parameters:
+  container: ucxx_rapidsai_ci_wheel
+  slices: []
+
+jobs:
+  - ${{ each slice in parameters.slices }}:
+    - job: ucxx_wheel_ucxx_build_${{ slice.name }}
+      workspace:
+        clean: all
+      pool:
+        name: MLNX
+        demands: ${{ slice.demands }}
+      displayName: 'UCXX wheel-build-ucxx (${{ slice.name }})'
+      container: ${{ parameters.container }}
+      timeoutInMinutes: 60
+      dependsOn: ucxx_wheel_libucxx_build_${{ slice.libucxx_slice }}
+      variables:
+        UCX_DIR: $(Agent.BuildDirectory)/ucx
+        UCXX_DIR: $(Agent.BuildDirectory)/ucxx
+        RAPIDS_CUDA_VERSION: ${{ slice.rapids_cuda_version }}
+        RAPIDS_PY_VERSION: ${{ slice.rapids_py_version }}
+        RAPIDS_BLD_OUTPUT_DIR: $(Build.ArtifactStagingDirectory)
+        WHEEL_INPUT_DIR: $(Build.ArtifactStagingDirectory)/wheel-libucxx
+
+      steps:
+        - checkout: self
+          path: ucx
+          fetchDepth: 100
+          retryCountOnTaskFailure: 5
+
+        - checkout: ucxx
+          path: ucxx
+          retryCountOnTaskFailure: 5
+
+        - task: DownloadBuildArtifacts@1
+          displayName: Fetch libucxx wheel artifact
+          inputs:
+            buildType: current
+            artifactName: ucxx-wheel-libucxx-${{ slice.libucxx_slice }}
+            downloadPath: $(System.DefaultWorkingDirectory)/_dl
+
+        - bash: |
+            rm -rf "$(WHEEL_INPUT_DIR)"
+            mv $(System.DefaultWorkingDirectory)/_dl/ucxx-wheel-libucxx-${{ slice.libucxx_slice }} "$(WHEEL_INPUT_DIR)"
+          displayName: Stage libucxx wheel artifact
+
+        - bash: bash $(UCX_DIR)/buildlib/tools/build_ucxx.sh wheel_ucxx
+          displayName: Build ucxx wheel
+
+        - task: PublishBuildArtifacts@1
+          displayName: Publish ucxx wheel artifact
+          inputs:
+            pathToPublish: $(RAPIDS_BLD_OUTPUT_DIR)
+            artifactName: ucxx-wheel-ucxx-${{ slice.name }}
diff --git a/buildlib/tools/build_ucxx.sh b/buildlib/tools/build_ucxx.sh
new file mode 100755
index 00000000000..193c6f8fdcd
--- /dev/null
+++ b/buildlib/tools/build_ucxx.sh
@@ -0,0 +1,80 @@
+#!/bin/bash -eE
+#
+# Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# See file LICENSE for terms.
+#
+# UCXX_build stage runner.
+# Usage: build_ucxx.sh <conda_cpp|conda_python|wheel_libucxx|wheel_ucxx|wheel_distributed_ucxx>
+# Env: RAPIDS_CUDA_VERSION, RAPIDS_PY_VERSION, UCXX_DIR, RAPIDS_BLD_OUTPUT_DIR
+
+phase=${1:?phase required}
+: "${RAPIDS_CUDA_VERSION:?RAPIDS_CUDA_VERSION required}"
+: "${RAPIDS_PY_VERSION:?RAPIDS_PY_VERSION required}"
+: "${UCXX_DIR:?UCXX_DIR required}"
+: "${RAPIDS_BLD_OUTPUT_DIR:?RAPIDS_BLD_OUTPUT_DIR required}"
+
+export RAPIDS_CUDA_VERSION RAPIDS_PY_VERSION
+mkdir -p "$RAPIDS_BLD_OUTPUT_DIR"
+
+# Upstream scripts read phase-specific output-dir env names.
+case "$phase" in
+  conda_*)  export RAPIDS_CONDA_BLD_OUTPUT_DIR="$RAPIDS_BLD_OUTPUT_DIR" ;;
+  wheel_*)  export RAPIDS_WHEEL_BLD_OUTPUT_DIR="$RAPIDS_BLD_OUTPUT_DIR" ;;
+esac
+
+# Shim RAPIDS gha-tools downloaders to local build output dir.
+mkdir -p "$HOME/.local/bin"
+for tool in rapids-download-conda-from-github rapids-download-from-github; do
+  printf '#!/bin/bash\necho "%s"\n' "$RAPIDS_BLD_OUTPUT_DIR" > "$HOME/.local/bin/$tool"
+  chmod +x "$HOME/.local/bin/$tool"
+done
+
+# sccache from rapids-configure-sccache crashes on CMake TryCompile in this
+# image; replace with a no-op that still exports the env vars rattler expects.
+cat > "$HOME/.local/bin/rapids-configure-sccache" <<'EOF'
+#!/bin/bash
+export CMAKE_C_COMPILER_LAUNCHER=
+export CMAKE_CXX_COMPILER_LAUNCHER=
+export CMAKE_CUDA_COMPILER_LAUNCHER=
+export RUSTC_WRAPPER=
+EOF
+chmod +x "$HOME/.local/bin/rapids-configure-sccache"
+
+# wheel_ucxx phase consumes a libucxx wheel artifact; shim downloader to it.
+if [ -n "${WHEEL_INPUT_DIR:-}" ]; then
+  printf '#!/bin/bash\necho "%s"\n' "$WHEEL_INPUT_DIR" > "$HOME/.local/bin/rapids-download-wheels-from-github"
+  chmod +x "$HOME/.local/bin/rapids-download-wheels-from-github"
+fi
+
+export PATH="$HOME/.local/bin:$PATH"
+
+cd "$UCXX_DIR"
+
+# Upstream header missing <unistd.h>.
+hdr=python/ucxx/ucxx/examples/python_future_task.h
+grep -q "include <unistd.h>" "$hdr" || sed -i '/^#pragma once/a #include <unistd.h>' "$hdr"
+
+echo "=== id: $(id), arch: $(uname -m), phase: $phase ==="
+
+# Wheel image (Rocky 8) defaults to gcc 8.5; UCXX C++ needs gcc-toolset-14
+# (designated initializers in assignment, matches upstream wheel CI).
+case "$phase" in
+  wheel_*)
+    toolset=/opt/rh/gcc-toolset-14/root/usr/bin
+    if [ -d "$toolset" ]; then
+      export PATH="$toolset:$PATH"
+      export CC="$toolset/gcc"
+      export CXX="$toolset/g++"
+    fi
+    ;;
+esac
+
+case "$phase" in
+  conda_cpp)             bash ci/build_cpp.sh ;;
+  conda_python)          bash ci/build_python.sh ;;
+  wheel_libucxx)         bash ci/build_wheel_libucxx.sh ;;
+  wheel_ucxx)            bash ci/build_wheel_ucxx.sh ;;
+  wheel_distributed_ucxx) bash ci/build_wheel_distributed_ucxx.sh ;;
+  *) echo "Unknown phase: $phase" >&2; exit 1 ;;
+esac
diff --git a/buildlib/tools/test_ucxx.sh b/buildlib/tools/test_ucxx.sh
new file mode 100755
index 00000000000..b4f0bf9588d
--- /dev/null
+++ b/buildlib/tools/test_ucxx.sh
@@ -0,0 +1,89 @@
+#!/bin/bash -eE
+#
+# Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# See file LICENSE for terms.
+#
+# UCXX_tests stage runner.
+# Usage: test_ucxx.sh <build|test_cpp|test_python>
+# Env: IS_GPU, RAPIDS_CUDA_VERSION, RAPIDS_PY_VERSION, UCXX_DIR
+
+phase=${1:?phase required: build | test_cpp | test_python}
+: "${IS_GPU:?IS_GPU required}"
+: "${RAPIDS_CUDA_VERSION:?RAPIDS_CUDA_VERSION required}"
+: "${RAPIDS_PY_VERSION:?RAPIDS_PY_VERSION required}"
+: "${UCXX_DIR:?UCXX_DIR required}"
+
+export RAPIDS_CUDA_VERSION RAPIDS_PY_VERSION
+export RAPIDS_CONDA_BLD_OUTPUT_DIR=/tmp/conda-bld-output
+mkdir -p "$RAPIDS_CONDA_BLD_OUTPUT_DIR"
+
+# Shim RAPIDS gha-tools downloaders to local conda-bld dir.
+mkdir -p "$HOME/.local/bin"
+for tool in rapids-download-conda-from-github rapids-download-from-github; do
+  printf '#!/bin/bash\necho "%s"\n' "$RAPIDS_CONDA_BLD_OUTPUT_DIR" > "$HOME/.local/bin/$tool"
+  chmod +x "$HOME/.local/bin/$tool"
+done
+export PATH="$HOME/.local/bin:$PATH"
+
+cd "$UCXX_DIR"
+
+case "$phase" in
+  build)
+    echo "=== id: $(id), arch: $(uname -m), gpu: $IS_GPU ==="
+    which python && python --version
+    [ "$IS_GPU" = "true" ] && nvidia-smi || true
+
+    if [ "$IS_GPU" = "true" ]; then
+      # sccache from rapids-configure-sccache crashes on CMake TryCompile in
+      # this image; replace with a no-op that still exports the env vars the
+      # rattler recipe expects.
+      cat > "$HOME/.local/bin/rapids-configure-sccache" <<'EOF'
+#!/bin/bash
+export CMAKE_C_COMPILER_LAUNCHER=
+export CMAKE_CXX_COMPILER_LAUNCHER=
+export CMAKE_CUDA_COMPILER_LAUNCHER=
+export RUSTC_WRAPPER=
+EOF
+      chmod +x "$HOME/.local/bin/rapids-configure-sccache"
+
+      # Upstream header missing <unistd.h>.
+      hdr=python/ucxx/ucxx/examples/python_future_task.h
+      grep -q "include <unistd.h>" "$hdr" || sed -i '/^#pragma once/a #include <unistd.h>' "$hdr"
+    else
+      # No GPU on CPU slice; soften nvidia-smi check.
+      sed -i 's#^  nvidia-smi$#  command -v nvidia-smi >/dev/null \&\& nvidia-smi || echo "(no GPU - CPU slice)"#' ci/test_common.sh
+    fi
+
+    bash ci/build_cpp.sh
+    bash ci/build_python.sh
+    ;;
+
+  test_cpp)
+    if [ "$IS_GPU" = "true" ]; then
+      # Run as Azure-injected step user (mapped to host swx-azure-svc uid),
+      # which matches the host MPS daemon owner so MPS accepts the client.
+      bash ci/test_cpp.sh
+    else
+      # CPU slice: no GPU driver loaded. Pin UCX onto host transports;
+      # filter CUDA-only gtest suites.
+      export CUDA_VISIBLE_DEVICES=
+      export UCX_TLS=tcp,sm,self
+      export GTEST_FILTER='-RMM*.*:CCCL*.*'
+      bash ci/test_cpp.sh
+    fi
+    ;;
+
+  test_python)
+    if [ "$IS_GPU" != "true" ]; then
+      echo "test_python only runs on GPU slice; skipping" >&2
+      exit 0
+    fi
+    bash ci/test_python.sh
+    ;;
+
+  *)
+    echo "Unknown phase: $phase (expected build | test_cpp | test_python)" >&2
+    exit 1
+    ;;
+esac