diff --git a/.github/workflows/gpu_ci_trigger.yml b/.github/workflows/gpu_ci_trigger.yml
index 7f7a8e7..80d4415 100644
--- a/.github/workflows/gpu_ci_trigger.yml
+++ b/.github/workflows/gpu_ci_trigger.yml
@@ -63,8 +63,10 @@ jobs:
           # 3. Add GitLab SSH remote
           git remote add gitlab git@gitlab.mpcdf.mpg.de:maxlin/cunumpy.git
 
-          # 4. Force push (This automatically starts the GitLab Pipeline)
-          git push -f gitlab "HEAD:refs/heads/$TARGET_BRANCH"
+          # 4. Force push (This automatically starts the GitLab Pipeline). The pushed branch
+          #    is gpu-test-*, so pass the target branch for the max-models dependencies.
+          git push -f -o ci.variable="MAX_MODELS_TARGET=${{ github.base_ref || github.ref_name }}" \
+            gitlab "HEAD:refs/heads/$TARGET_BRANCH"
           echo "PUSHED_SHA=$(git rev-parse HEAD)" >> "$GITHUB_ENV"
 
           # 5. Provide the direct link
diff --git a/.github/workflows/macos.yml b/.github/workflows/macos.yml
index b13bab3..388d28f 100644
--- a/.github/workflows/macos.yml
+++ b/.github/workflows/macos.yml
@@ -43,6 +43,8 @@ jobs:
       - name: Install project
         run: |
           pip install --upgrade pip
+          # devel branches of the max-models dependencies, except on runs that target main
+          if [ "${GITHUB_BASE_REF:-$GITHUB_REF_NAME}" != main ]; then python -m pip install -r requirements-devel.txt; fi
           pip install ".[test-compiled,metal]"
 
       # Hosted macOS runners are virtual machines and usually have no Metal GPU,
diff --git a/.github/workflows/pyodide.yml b/.github/workflows/pyodide.yml
index dc35744..c431dee 100644
--- a/.github/workflows/pyodide.yml
+++ b/.github/workflows/pyodide.yml
@@ -6,6 +6,8 @@ on:
   pull_request:
     branches: [main, devel]
 
+# PyPI only, also on devel: micropip, which installs the wheel's dependencies in Pyodide,
+# cannot install from git.
 jobs:
   wheel-test:
     runs-on: ubuntu-latest
diff --git a/.github/workflows/static_analysis.yml b/.github/workflows/static_analysis.yml
index 02b58a7..ae0eb4d 100644
--- a/.github/workflows/static_analysis.yml
+++ b/.github/workflows/static_analysis.yml
@@ -82,6 +82,7 @@ jobs:
 
       - name: Docstring sections with numpydoc
         run: |
+          if [ "${GITHUB_BASE_REF:-$GITHUB_REF_NAME}" != main ]; then python -m pip install -r requirements-devel.txt; fi
           pip install . numpydoc
           python -m numpydoc lint $(git ls-files 'src/cunumpy/*.py' 'src/cunumpy/**/*.py' | grep -v _fake_cupy)
 
diff --git a/.github/workflows/testing.yml b/.github/workflows/testing.yml
index 0f4d562..1676664 100644
--- a/.github/workflows/testing.yml
+++ b/.github/workflows/testing.yml
@@ -10,6 +10,8 @@ on:
       - devel
   workflow_call:
 
+# Runs that do not target main install the devel branches of the max-models dependencies
+# (requirements-devel.txt) before cunumpy; PRs to main and the release use PyPI only.
 jobs:
   compiled-fake-cupy:
     runs-on: ubuntu-latest
@@ -22,6 +24,7 @@ jobs:
         run: |
           sudo apt-get update
           sudo apt-get install -y gcc gfortran
+          if [ "${GITHUB_BASE_REF:-$GITHUB_REF_NAME}" != main ]; then python -m pip install -r requirements-devel.txt; fi
           python -m pip install ".[test-compiled]"
       - name: Test compiled host kernels with fake device arrays
         env:
@@ -61,6 +64,7 @@ jobs:
       - name: Install project
         run: |
           pip install --upgrade pip
+          if [ "${GITHUB_BASE_REF:-$GITHUB_REF_NAME}" != main ]; then python -m pip install -r requirements-devel.txt; fi
           pip install ".[dev]"
 
       - name: Run tests
@@ -70,3 +74,18 @@ jobs:
       - name: Run docstring examples
         run: |
           pytest --doctest-modules src/cunumpy
+
+  # PyPI rejects direct references (`name @ git+...`): the wheel's metadata must have none.
+  no-direct-references:
+    if: (github.base_ref || github.ref_name) == 'main'
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - name: Check the wheel for direct references
+        run: |
+          python -m pip install build
+          python -m build --wheel
+          ! unzip -p dist/*.whl '*.dist-info/METADATA' | grep -E '^Requires-Dist:.* @ '
diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
index 1d852e4..cba3e93 100644
--- a/.gitlab-ci.yml
+++ b/.gitlab-ci.yml
@@ -14,6 +14,13 @@ variables:
 stages:
   - test
 
+# Runs that do not target main install the devel branches of the max-models dependencies
+# first; main uses PyPI only. The GitHub trigger pushes gpu-test-* branches and passes the
+# target branch in MAX_MODELS_TARGET.
+.install_devel_dependencies: &install_devel_dependencies
+  - TARGET="${MAX_MODELS_TARGET:-${CI_MERGE_REQUEST_TARGET_BRANCH_NAME:-$CI_COMMIT_BRANCH}}"
+  - if [ "$TARGET" != main ]; then python3 -m pip install --user -r requirements-devel.txt; fi
+
 .gpu_job_nvidia:
   stage: test
   image: ${CUDA_IMAGE}
@@ -27,6 +34,7 @@ stages:
     - nvidia-smi
     # Keep NVRTC, CUDA libraries and headers on the same toolkit generation.
     - python3 -m pip install --user "cupy-cuda13x[ctk]${CUPY_PIN}" "cuda-toolkit==13.2.*"
+    - *install_devel_dependencies
     - python3 -m pip install --user -e ".[test]"
     - export PATH="$HOME/.local/bin:$PATH"
     - python3 -c "import cupy; cupy.show_config()"
@@ -49,6 +57,7 @@ stages:
     - export ROCM_HOME="${ROCM_PATH:-/opt/rocm}"
     - export CUPY_INSTALL_USE_HIP=1
     - python3 -m pip install --user "cupy${CUPY_PIN}"
+    - *install_devel_dependencies
     - python3 -m pip install --user -e ".[test]"
     - export PATH="$HOME/.local/bin:$PATH"
     - python3 -c "import cupy; cupy.show_config()"
diff --git a/docs/src/content/docs/getting-started/installation.md b/docs/src/content/docs/getting-started/installation.md
index 5a04447..7653a25 100644
--- a/docs/src/content/docs/getting-started/installation.md
+++ b/docs/src/content/docs/getting-started/installation.md
@@ -88,6 +88,11 @@ python -m pytest tests/unit
 
 Tests that need a GPU are skipped automatically where CuPy is not functional.
 
+On `devel`, cunumpy is developed against the `devel` branches of the other max-models packages
+(currently maybempi): `uv sync` picks them up from `[tool.uv.sources]`, and with pip run
+`python -m pip install -r requirements-devel.txt` before the install above. CI does the same for
+pull requests to `devel`; pull requests to `main` and the PyPI release use PyPI versions only.
+
 ## Environment variables
 
 | Variable | Effect |
diff --git a/pyproject.toml b/pyproject.toml
index f70ac73..34c15e3 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -47,6 +47,11 @@ where = [ "src" ]
 [tool.setuptools.package-data]
 cunumpy = [ "py.typed", "*.pyi", "LLM_GUIDE.md", "cuda/include/cunumpy/*.cuh" ]
 
+# PRs to devel test against the devel branches of the max-models packages; PRs to main
+# and the release use PyPI only.
+[tool.uv.sources]
+maybempi = { git = "https://github.com/max-models/maybempi", branch = "devel" }
+
 [tool.ruff]
 # agent worktrees and scratch scripts are local copies, not part of the project
 extend-exclude = [ ".claude" ]
diff --git a/requirements-devel.txt b/requirements-devel.txt
new file mode 100644
index 0000000..ab53722
--- /dev/null
+++ b/requirements-devel.txt
@@ -0,0 +1,4 @@
+# The devel branches of the max-models packages in cunumpy's dependency tree. CI installs
+# them before cunumpy on every run that does not target main; main and the release use PyPI.
+#   pip install -r requirements-devel.txt
+maybempi @ git+https://github.com/max-models/maybempi@devel