diff --git a/CHANGELOG.md b/CHANGELOG.md
index fdcd5e3b..15aa0f93 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -17,6 +17,13 @@ changelog](https://keepachangelog.com/en/1.1.0/) format. This project follows
 ### Removed
 -->
 
+### Added
+
+- Experimental ASE neighbour-list plugin (`vesin._ase_plugin`): registers Vesin
+  under the `ase.plugins` entry point as a host `NeighborListFunction` backend,
+  plus an optional device-resident `DeviceNeighborList` capability (CUDA via
+  CuPy) for ASE's experimental device neighbour-list protocol.
+
 ### Changed
 
 - Importing from `vesin.torch` is deprecated, users should now import from
diff --git a/python/vesin/pyproject.toml b/python/vesin/pyproject.toml
index 7f9f5882..b4fcb6ab 100644
--- a/python/vesin/pyproject.toml
+++ b/python/vesin/pyproject.toml
@@ -32,6 +32,11 @@ classifiers = [
 
 [project.optional-dependencies]
 torch = ["vesin-torch"]
+ase = ["ase"]
+
+# ASE v4 neighbour-list plugin discovery (registers nothing on older ASE).
+[project.entry-points."ase.plugins"]
+vesin = "vesin._ase_plugin"
 
 [project.urls]
 homepage = "https://github.com/Luthaf/vesin/"
diff --git a/python/vesin/tests/test_ase_plugin.py b/python/vesin/tests/test_ase_plugin.py
new file mode 100644
index 00000000..468b1c07
--- /dev/null
+++ b/python/vesin/tests/test_ase_plugin.py
@@ -0,0 +1,174 @@
+import ase.build
+import ase.neighborlist
+import numpy as np
+import pytest
+
+from vesin._ase_plugin import device_neighbor_list, neighbor_list
+
+
+def _ijS_set(i, j, S):
+    i = np.asarray(i, dtype=np.int64)
+    j = np.asarray(j, dtype=np.int64)
+    S = np.asarray(S, dtype=np.int64)
+    return set(
+        zip(
+            i.tolist(),
+            j.tolist(),
+            *[S[:, k].tolist() for k in range(3)],
+            strict=True,
+        )
+    )
+
+
+def _system():
+    atoms = ase.build.bulk("Cu", "fcc", a=3.6, cubic=True).repeat((2, 2, 2))
+    return atoms
+
+
+# --------------------------------------------------------------------------- #
+# Host plugin (no GPU)
+# --------------------------------------------------------------------------- #
+def test_host_matches_ase():
+    atoms = _system()
+    cutoff = 4.0
+    vi, vj, vd, vD, vS = neighbor_list("ijdDS", atoms, cutoff)
+    ai, aj, ad, aD, aS = ase.neighborlist.neighbor_list("ijdDS", atoms, cutoff)
+    assert len(vi) == len(ai)
+    assert _ijS_set(vi, vj, vS) == _ijS_set(ai, aj, aS)
+
+
+def test_host_self_interaction_rejected():
+    atoms = _system()
+    with pytest.raises(NotImplementedError):
+        neighbor_list("ij", atoms, 4.0, self_interaction=True)
+
+
+def test_plugin_registration():
+    # Exposes __ase_plugins__; the "vesin" plugin is present iff ASE has the v4
+    # plugin API (older ASE -> empty set, by design).
+    from vesin import _ase_plugin
+
+    try:
+        from ase._4.plugins.neighborlist import NeighborListPlugin  # noqa: F401
+    except ImportError:
+        assert _ase_plugin.__ase_plugins__ == set()
+    else:
+        names = {p.name for p in _ase_plugin.__ase_plugins__}
+        assert "vesin" in names
+
+
+# --------------------------------------------------------------------------- #
+# Device backend (CUDA via CuPy + ASE v4 device protocol)
+# --------------------------------------------------------------------------- #
+def _device_skip_reason():
+    try:
+        import cupy as cp
+    except ImportError as exc:
+        return f"cupy not available: {exc}"
+    try:
+        cp.cuda.Device(0).compute_capability
+    except Exception as exc:  # pragma: no cover - env dependent
+        return f"CUDA not available: {exc}"
+    try:
+        import ase._4.plugins.neighborlist_device  # noqa: F401
+    except ImportError:
+        return "ASE has no experimental device neighbour-list protocol"
+    return None
+
+
+_DEVICE_SKIP = _device_skip_reason()
+device = pytest.mark.skipif(_DEVICE_SKIP is not None, reason=str(_DEVICE_SKIP))
+
+
+@device
+def test_device_satisfies_protocol():
+    from ase._4.plugins.neighborlist_device import DeviceNeighborList
+
+    be = device_neighbor_list()
+    assert isinstance(be, DeviceNeighborList)
+    assert be.differentiable is False
+    assert be.device[0] == 2  # CUDA
+
+
+@device
+def test_device_equivalence_vs_ase():
+    import cupy as cp
+
+    atoms = _system()
+    cutoff = 4.0
+    res = device_neighbor_list().build_device(
+        cp.asarray(atoms.positions),
+        np.asarray(atoms.cell[:]),
+        tuple(bool(b) for b in atoms.pbc),
+        cutoff,
+        "ijSD",
+    )
+    vi = cp.asnumpy(res.get("i"))
+    vj = cp.asnumpy(res.get("j"))
+    vS = cp.asnumpy(res.get("S"))
+    assert res.get("i").__dlpack_device__()[0] == 2  # device-resident
+
+    ai, aj, aS, aD = ase.neighborlist.neighbor_list("ijSD", atoms, cutoff)
+    assert _ijS_set(vi, vj, vS) == _ijS_set(ai, aj, aS)
+
+
+@device
+def test_device_needs_rebuild_delegates():
+    # Reuse is delegated to Vesin's persistent NeighborList(skin=); needs_rebuild
+    # is a no-op device-resident True (the consumer always calls build_device).
+    import cupy as cp
+
+    atoms = _system()
+    be = device_neighbor_list()
+    be.build_device(
+        cp.asarray(atoms.positions),
+        np.asarray(atoms.cell[:]),
+        tuple(bool(b) for b in atoms.pbc),
+        4.0,
+        "ijS",
+    )
+    nr = be.needs_rebuild(cp.asarray(atoms.positions), skin=0.4)
+    assert nr.__dlpack_device__()[0] == 2  # device-resident
+    assert bool(nr) is True
+
+
+@device
+def test_device_skin_reuse_correct():
+    # With skin>0 the persistent calculator reuses the list across a sub-skin
+    # move; the result must still match a fresh build (Vesin filters to the true
+    # cutoff on reuse).
+    import cupy as cp
+
+    from vesin._ase_device import VesinDeviceNeighborList
+
+    atoms = _system()
+    cell = np.asarray(atoms.cell[:])
+    pbc = tuple(bool(b) for b in atoms.pbc)
+    be = VesinDeviceNeighborList(skin=0.5)
+    be.build_device(cp.asarray(atoms.positions), cell, pbc, 4.0, "ijS")  # seed cache
+    moved = atoms.positions + 0.05  # sub-skin displacement -> reuse path
+    res = be.build_device(cp.asarray(moved), cell, pbc, 4.0, "ijS")
+    vi = cp.asnumpy(res.get("i"))
+    vj = cp.asnumpy(res.get("j"))
+    vS = cp.asnumpy(res.get("S"))
+
+    a2 = atoms.copy()
+    a2.positions = moved
+    ai, aj, aS = ase.neighborlist.neighbor_list("ijS", a2, 4.0)
+    assert _ijS_set(vi, vj, vS) == _ijS_set(ai, aj, aS)
+
+
+@device
+def test_device_padded_unsupported():
+    import cupy as cp
+
+    atoms = _system()
+    be = device_neighbor_list()
+    with pytest.raises(NotImplementedError):
+        be.build_device(
+            cp.asarray(atoms.positions),
+            np.asarray(atoms.cell[:]),
+            tuple(bool(b) for b in atoms.pbc),
+            4.0,
+            max_capacity=64,
+        )
diff --git a/python/vesin/vesin/_ase_device.py b/python/vesin/vesin/_ase_device.py
new file mode 100644
index 00000000..a49b8342
--- /dev/null
+++ b/python/vesin/vesin/_ase_device.py
@@ -0,0 +1,181 @@
+"""Experimental: device-resident neighbour-list adapter for ASE.
+
+Adapts Vesin's CUDA cell list (reached through its CuPy array interface) to ASE's
+experimental device neighbour-list protocol
+(:mod:`ase._4.plugins.neighborlist_device`): :class:`DeviceNeighborList` and
+:class:`DeviceNeighborResult`. A device-resident calculator discovers it via
+``isinstance(backend, DeviceNeighborList)`` and exchanges edge data on-device via
+DLPack, with no host round-trip per step.
+
+GPU-only and requires CuPy: passing CuPy device positions to
+:meth:`vesin.NeighborList.compute` dispatches to the CUDA kernel and returns
+device-resident CuPy arrays. The host path is unchanged (``vesin.ase_neighbor_list``).
+
+Notes / limitations
+-------------------
+* ``differentiable = False`` here only because this adapter goes through the
+  **CuPy** path, which is not an autograd framework -- not a Vesin limitation.
+  Vesin's torch binding (``vesin-torch``) runs on GPU and is autograd-
+  differentiable, so it would make a natural ``differentiable = True`` device
+  backend for this protocol (a useful follow-up).
+* No fixed-capacity / dense output -- Vesin currently returns COO only, so the
+  padded (``max_capacity``) path is unsupported and raises. (Dense output is in
+  progress upstream; the padded path can map onto it once available.)
+* Reuse is delegated to Vesin: the adapter holds a **persistent**
+  ``NeighborList(skin=...)`` and ``build_device`` calls ``.compute`` on it, so
+  Vesin's own Verlet logic reuses the list across steps (displacement, and box
+  changes once Luthaf/vesin#172 lands -- NPT). ``needs_rebuild`` therefore just
+  returns a device-resident ``True`` (a no-op signal): the consumer always calls
+  ``build_device`` and Vesin makes the reuse cheap internally. The Verlet skin is
+  set at construction (``VesinDeviceNeighborList(skin=...)``), matching Vesin's
+  API; ``needs_rebuild``'s ``skin`` argument is unused. (A compiled consumer that
+  wants to *skip* the build itself would instead need an explicit device-scalar
+  displacement check -- an open ASE-protocol design point.)
+* Scalar cutoff only; ``self_interaction=True`` rejected -- both raise.
+"""
+
+import numpy as np
+
+
+def _cupy():
+    import cupy as cp  # imported lazily; this adapter is GPU-only
+
+    return cp
+
+
+def _to_device_cell(cell):
+    """Return the (3, 3) cell as a CuPy array (Vesin needs box and points to be
+    the same array type). The cell is tiny, so any host->device copy here is not
+    the residency concern (that is the O(n_atoms) positions / O(n_edges) edges)."""
+    cp = _cupy()
+    if isinstance(cell, cp.ndarray):
+        return cell
+    to_dlpack_device = getattr(cell, "__dlpack_device__", None)
+    if to_dlpack_device is not None and to_dlpack_device()[0] != 1:  # device array
+        return cp.from_dlpack(cell)
+    return cp.asarray(np.ascontiguousarray(np.asarray(cell), dtype=float))
+
+
+class VesinDeviceResult:
+    """On-device neighbour data (CuPy); satisfies ``DeviceNeighborResult``.
+
+    Vesin returns COO arrays only, so this result is never padded.
+    """
+
+    def __init__(self, arrays, *, n_edges):
+        self._arrays = arrays  # name -> CuPy device array (DLPack-exporting)
+        self._n_edges = n_edges
+
+    @property
+    def n_edges(self):
+        return int(self._n_edges)
+
+    @property
+    def did_overflow(self):
+        return False
+
+    @property
+    def padded(self):
+        return False
+
+    def get(self, quantity):
+        try:
+            return self._arrays[quantity]
+        except KeyError:
+            available = ", ".join(sorted(self._arrays)) or "(none)"
+            raise KeyError(
+                f"quantity {quantity!r} not available; have {available}. "
+                "It was not among the requested quantities."
+            ) from None
+
+    def mask(self):
+        return None
+
+
+class VesinDeviceNeighborList:
+    """Vesin device backend; satisfies ``DeviceNeighborList``."""
+
+    differentiable = False
+
+    def __init__(self, device_id=0, skin=0.0):
+        self._device_type = 2  # kDLCUDA (Vesin's GPU backend is CUDA)
+        self._device_id = int(device_id)
+        # Persistent calculator so Vesin's own Verlet ``skin`` reuse kicks in
+        # across steps. skin=0 means rebuild every call (no reuse).
+        self._skin = float(skin)
+        self._nl = None
+        self._nl_cutoff = None
+
+    @property
+    def device(self):
+        return (self._device_type, self._device_id)
+
+    def build_device(
+        self,
+        positions,
+        cell,
+        pbc,
+        cutoff,
+        quantities="ijS",
+        *,
+        self_interaction=False,
+        max_capacity=None,
+        stream=None,
+    ):
+        if self_interaction:
+            raise NotImplementedError(
+                "the Vesin backend does not support self_interaction=True"
+            )
+        if isinstance(cutoff, dict) or np.ndim(cutoff) != 0:
+            raise NotImplementedError(
+                "the Vesin device backend supports only a scalar cutoff"
+            )
+        if max_capacity is not None:
+            raise NotImplementedError(
+                "Vesin has no fixed-capacity (dense) output; use the tight path "
+                "(max_capacity=None) for COO i/j/S/D"
+            )
+        invalid = set(quantities) - set("ijdDS")
+        if invalid or not quantities:
+            raise ValueError(
+                f'quantities must be a non-empty subset of "ijdDS"; got {quantities!r}.'
+            )
+
+        from ._neighbors import NeighborList
+
+        cp = _cupy()
+        pos = cp.from_dlpack(positions)  # device array (zero-copy view)
+        self._device_id = int(pos.device.id)
+        box = _to_device_cell(cell)
+        pbc_t = tuple(bool(b) for b in pbc)
+
+        # Reuse one persistent NeighborList so Vesin's Verlet ``skin`` reuse spans
+        # calls (rebuilt only if the cutoff changes). Vesin decides internally
+        # whether to rebuild or reuse on each ``.compute``.
+        if self._nl is None or self._nl_cutoff != float(cutoff):
+            self._nl = NeighborList(
+                cutoff=float(cutoff), full_list=True, skin=self._skin
+            )
+            self._nl_cutoff = float(cutoff)
+        out = self._nl.compute(pos, box, pbc_t, quantities=quantities)
+        if not isinstance(out, (list, tuple)):  # single-quantity -> bare array
+            out = (out,)
+        # Keyed by NAME (Vesin returns in requested order; we never rely on it).
+        arrays = {name: arr for name, arr in zip(quantities, out, strict=True)}
+        n_edges = int(arrays[quantities[0]].shape[0])
+        return VesinDeviceResult(arrays, n_edges=n_edges)
+
+    def needs_rebuild(self, positions, *, skin, stream=None):
+        """Delegate reuse to Vesin: return a device-resident ``True`` scalar.
+
+        Vesin's persistent ``NeighborList(skin=...)`` owns the rebuild-vs-reuse
+        decision internally (displacement, and box changes once Luthaf/vesin#172
+        lands), so the adapter always signals "build" and lets ``build_device``'s
+        persistent calculator reuse the list cheaply. The ``skin`` argument is not
+        used here -- the Verlet skin is set at construction
+        (``VesinDeviceNeighborList(skin=...)``), matching Vesin's API. (A compiled
+        consumer that needs to *skip* the build itself would instead want an
+        explicit device-scalar displacement check; see the ASE protocol notes.)
+        """
+        cp = _cupy()
+        return cp.asarray(True)
diff --git a/python/vesin/vesin/_ase_plugin.py b/python/vesin/vesin/_ase_plugin.py
new file mode 100644
index 00000000..26086741
--- /dev/null
+++ b/python/vesin/vesin/_ase_plugin.py
@@ -0,0 +1,69 @@
+"""ASE neighbour-list plugin for Vesin (host + experimental device capability).
+
+Registers Vesin as an ``ase.plugins`` neighbour-list backend so it can be
+selected via :func:`ase.neighborlist.get_neighbor_list` (ASE v4 plugin API).
+Selection is never automatic -- this only makes the backend *available* under the
+name ``"vesin"``.
+
+This module is the ``ase.plugins`` entry point (it exposes ``__ase_plugins__``).
+Registration is guarded so installing Vesin alongside an ASE that predates the v4
+plugin API registers nothing rather than breaking plugin discovery. The heavy
+imports (ase, CuPy) happen lazily, inside the adapters.
+"""
+
+from __future__ import annotations
+
+
+def neighbor_list(quantities, atoms, cutoff, *, self_interaction=False):
+    """Vesin neighbour list adapted to ASE's ``NeighborListFunction`` protocol.
+
+    Thin wrapper over :func:`vesin.ase_neighbor_list`, returning the same
+    ``(i, j, d, D, S)`` flat arrays and quantity letters as
+    :func:`ase.neighborlist.neighbor_list`. Vesin has no ``self_interaction``
+    option (it never returns pure self-pairs); per the plugin contract a request
+    it cannot honour is rejected rather than silently differing.
+    """
+    if self_interaction:
+        raise NotImplementedError(
+            "the vesin backend does not support self_interaction=True"
+        )
+    from ._ase import ase_neighbor_list
+
+    return ase_neighbor_list(quantities, atoms, cutoff)
+
+
+def device_neighbor_list(device_id=0, skin=0.0):
+    """Return the *experimental* device-resident Vesin backend.
+
+    The result satisfies ASE's experimental ``DeviceNeighborList`` protocol
+    (:mod:`ase._4.plugins.neighborlist_device`) and builds edge data on-device
+    (exchanged via DLPack). GPU-only (CUDA) and requires CuPy.
+
+    ``skin`` sets the Verlet skin of the backend's persistent ``NeighborList`` so
+    it reuses the list across steps (``skin=0`` rebuilds every call).
+
+    The device capability is separate from the host ``neighbor_list``
+    registration: the host ``NeighborListPlugin`` carries no device slot, so a
+    consumer obtains the device backend through this factory and checks
+    ``isinstance(backend, DeviceNeighborList)``.
+    """
+    from ._ase_device import VesinDeviceNeighborList
+
+    return VesinDeviceNeighborList(device_id=device_id, skin=skin)
+
+
+try:
+    from ase._4.plugins.neighborlist import NeighborListPlugin
+except ImportError:
+    # ASE without the v4 plugin API: register nothing, but do not break the
+    # discovery of other plugins.
+    __ase_plugins__: set = set()
+else:
+    __ase_plugins__ = {
+        NeighborListPlugin(
+            "vesin",
+            long_name="Vesin neighbour list",
+            citation="https://github.com/Luthaf/vesin",
+            implementation="vesin._ase_plugin.neighbor_list",
+        ),
+    }