diff --git a/requirements/CI-tests-complete/requirements.txt b/requirements/CI-tests-complete/requirements.txt
index 5c859cf8..b590814d 100644
--- a/requirements/CI-tests-complete/requirements.txt
+++ b/requirements/CI-tests-complete/requirements.txt
@@ -16,8 +16,9 @@ pytest-cov==3.0.0
 pytest-xdist==2.5.0
 seaborn==0.11.2
 setuptools==65.4.1
+sgkit[vcf]==0.5.0
 sortedcontainers==2.4.0
 tqdm==4.64.0
 tskit==0.5.3
 twine==4.0.1
-zarr==2.11.3
+zarr==2.10.3
diff --git a/requirements/CI-tests-pip/requirements.txt b/requirements/CI-tests-pip/requirements.txt
index bda92366..a26320bb 100644
--- a/requirements/CI-tests-pip/requirements.txt
+++ b/requirements/CI-tests-pip/requirements.txt
@@ -8,7 +8,8 @@ pandas==1.4.2
 pytest==7.1.2
 pytest-xdist==2.5.0
 seaborn==0.11.2
+sgkit[vcf]==0.5.0
 sortedcontainers==2.4.0
 tqdm==4.64.0
 tskit==0.5.3
-zarr==2.11.3
+zarr==2.10.3
diff --git a/requirements/CI-tests-pip/requirements3.7.txt b/requirements/CI-tests-pip/requirements3.7.txt
index 55c17a53..52917d58 100644
--- a/requirements/CI-tests-pip/requirements3.7.txt
+++ b/requirements/CI-tests-pip/requirements3.7.txt
@@ -8,7 +8,8 @@ pandas==1.3.5
 pytest==7.1.2
 pytest-xdist==2.5.0
 seaborn==0.11.2
+sgkit[vcf]==0.5.0
 sortedcontainers==2.4.0
 tqdm==4.64.0
 tskit==0.5.3
-zarr==2.11.3
+zarr==2.10.3
diff --git a/requirements/development.txt b/requirements/development.txt
index 52523426..9a2bf29e 100644
--- a/requirements/development.txt
+++ b/requirements/development.txt
@@ -33,4 +33,4 @@ pandas
 matplotlib
 seaborn
 colorama
-
+sgkit[vcf]
diff --git a/tests/test_sgkit.py b/tests/test_sgkit.py
new file mode 100644
index 00000000..1141f838
--- /dev/null
+++ b/tests/test_sgkit.py
@@ -0,0 +1,45 @@
+#
+# Copyright (C) 2022 University of Oxford
+#
+# This file is part of tsinfer.
+#
+# tsinfer is free software: you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# tsinfer is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with tsinfer.  If not, see <http://www.gnu.org/licenses/>.
+#
+"""
+Tests for the data files.
+"""
+import msprime
+import numpy as np
+import sgkit.io.vcf
+
+import tsinfer
+
+
+def test_sgkit_dataset(tmp_path):
+    ts = msprime.sim_ancestry(
+        samples=50,
+        ploidy=3,
+        recombination_rate=0.25,
+        sequence_length=50,
+        random_seed=100,
+    )
+    ts = msprime.sim_mutations(ts, rate=0.025, model=msprime.BinaryMutationModel())
+    with open(tmp_path / "data.vcf", "w") as f:
+        ts.write_vcf(f)
+    sgkit.io.vcf.vcf_to_zarr(
+        tmp_path / "data.vcf", tmp_path / "data.zarr", ploidy=3, max_alt_alleles=1
+    )
+    samples = tsinfer.SgkitSampleData(tmp_path / "data.zarr")
+    inf_ts = tsinfer.infer(samples)
+    assert np.array_equal(ts.genotype_matrix(), inf_ts.genotype_matrix())
diff --git a/tsinfer/formats.py b/tsinfer/formats.py
index 7fcc7ceb..10f939fe 100644
--- a/tsinfer/formats.py
+++ b/tsinfer/formats.py
@@ -2100,7 +2100,7 @@ def variants(self, sites=None, recode_ancestral=None):
                 genos = geno_map[genos]
             yield Variant(site=site, alleles=alleles, genotypes=genos)
 
-    def __all_haplotypes(self, sites=None, recode_ancestral=None):
+    def _all_haplotypes(self, sites=None, recode_ancestral=None):
         # We iterate over chunks vertically here, and it's not worth complicating
         # the chunk iterator to handle this.
         if recode_ancestral is None:
@@ -2156,7 +2156,7 @@ def haplotypes(self, samples=None, sites=None, recode_ancestral=None):
                 raise ValueError("Sample index too large.")
         j = 0
 
-        for index, a in self.__all_haplotypes(sites, recode_ancestral):
+        for index, a in self._all_haplotypes(sites, recode_ancestral):
             if j == len(samples):
                 break
             if index == samples[j]:
@@ -2225,6 +2225,263 @@ def populations(self):
             yield Population(j, metadata=metadata)
 
 
+class SgkitSampleData(SampleData):
+
+    FORMAT_NAME = "tsinfer-sgkit-sample-data"
+    FORMAT_VERSION = (0, 1)
+
+    def __init__(self, path):
+        self.path = path
+        self.data = zarr.open(path, mode="r")
+        self._num_sites, self._num_individuals, self.ploidy = self.data[
+            "call_genotype"
+        ].shape
+        self._num_samples = self._num_individuals * self.ploidy
+
+        assert self.ploidy == self.data["call_genotype"].chunks[2]
+
+    def __metadata_schema_getter(self, zarr_group):
+        try:
+            return self.data[zarr_group].attrs["metadata_schema"]
+        except KeyError:
+            return {"codec": "json"}
+
+    @property
+    def uuid(self):
+        return (
+            "Hmm, not sure, could just generate a UUID, but then it wouldn't"
+            "be in the file - maybe we do need to write back on init"
+        )
+
+    @property
+    def format_name(self):
+        return self.FORMAT_NAME
+
+    @property
+    def format_version(self):
+        return self.FORMAT_VERSION
+
+    @property
+    def finalised(self):
+        return True
+
+    @property
+    def sequence_length(self):
+        try:
+            return self.data.attrs["sequence_length"]
+        except KeyError:
+            return int(np.max(self.data["variant_position"])) + 1
+
+    @property
+    def num_sites(self):
+        return self._num_sites
+
+    @property
+    def sites_metadata_schema(self):
+        return self.__metadata_schema_getter("sites")
+
+    @property
+    def sites_metadata(self):
+        try:
+            return self.data["sites/metadata"]
+        except KeyError:
+            return zarr.array(
+                [{}] * self.num_individuals, object_codec=numcodecs.JSON()
+            )
+
+    @property
+    def sites_time(self):
+        try:
+            return self.data["sites/time"]
+        except KeyError:
+            return np.full(self.data["variant_position"].shape, tskit.UNKNOWN_TIME)
+
+    @property
+    def sites_position(self):
+        return self.data["variant_position"]
+
+    @property
+    def sites_alleles(self):
+        return self.data["variant_allele"]
+
+    @property
+    def sites_ancestral_allele(self):
+        try:
+            return self.data["sites/ancestral_allele"]
+        except KeyError:
+            # Maintains backwards compatibility: in previous tsinfer versions the
+            # ancestral allele was always the zeroth element in the alleles list
+            return np.zeros(self.num_sites, dtype=np.int8)
+
+    @property
+    def sites_genotypes(self):
+        gt = self.data["call_genotype"]
+        return gt[...].reshape(gt.shape[0], gt.shape[1] * gt.shape[2])
+
+    @property
+    def provenances_timestamp(self):
+        try:
+            return self.data["provenances_timestamp"]
+        except KeyError:
+            return np.array([], dtype=object)
+
+    @property
+    def provenances_record(self):
+        try:
+            return self.data["provenances_record"]
+        except KeyError:
+            return np.array([], dtype=object)
+
+    @property
+    def num_samples(self):
+        return self._num_samples
+
+    @property
+    def samples_individual(self):
+        ret = np.zeros((self.num_samples), dtype=np.int32)
+        for p in range(self.ploidy):
+            ret[p :: self.ploidy] = np.arange(self.num_individuals)
+        return ret
+
+    @property
+    def metadata_schema(self):
+        try:
+            return self.data.attrs["metadata_schema"]
+        except KeyError:
+            None
+
+    @property
+    def metadata(self):
+        try:
+            return self.data.attrs["metadata_schema"]
+        except KeyError:
+            return b""
+
+    @property
+    def populations_metadata(self):
+        try:
+            return self.data["populations/metadata"]
+        except KeyError:
+            return np.array([], dtype=object)
+
+    @property
+    def populations_metadata_schema(self):
+        return self.__metadata_schema_getter("populations")
+
+    @property
+    def num_individuals(self):
+        return self._num_individuals
+
+    @property
+    def individuals_time(self):
+        try:
+            return self.data["individuals/time"]
+        except KeyError:
+            return np.full(self.num_individuals, tskit.UNKNOWN_TIME)
+
+    @property
+    def individuals_metadata_schema(self):
+        return self.__metadata_schema_getter("individuals")
+
+    @property
+    def individuals_metadata(self):
+        try:
+            return self.data["individuals/metadata"]
+        except KeyError:
+            return zarr.array(
+                [{}] * self.num_individuals, object_codec=numcodecs.JSON()
+            )
+
+    @property
+    def individuals_location(self):
+        try:
+            return self.data["individuals/location"]
+        except KeyError:
+            return zarr.array([[]] * self.num_individuals, dtype=float)
+
+    @property
+    def individuals_population(self):
+        try:
+            return self.data["individuals/population"]
+        except KeyError:
+            return np.full((self.num_individuals), tskit.NULL, dtype=np.int32)
+
+    @property
+    def individuals_flags(self):
+        try:
+            return self.data["individuals/population"]
+        except KeyError:
+            return np.full((self.num_individuals), 0, dtype=np.int32)
+
+    def variants(self, sites=None, recode_ancestral=None):
+        """
+        Returns an iterator over the :class:`Variant` objects. This is equivalent to
+        the :meth:`tskit.TreeSequence.variants` iterator. If recode_ancestral is
+        ``True``, the ``.alleles`` attribute of each variant is guaranteed to return
+        the alleles in an order such that the ancestral state is the first item
+        in the list. In this case, ``variant.alleles`` may list the alleles in a
+        different order from the input order as listed in ``variant.site.alleles``,
+        and the values in genotypes array will be recoded so that the ancestral
+        state will have a genotype of 0. If the ancestral state is unknown, the
+        original input order is kept.
+
+        :param array sites: A numpy array of ascending site ids for which to return
+            data. If None (default) return all sites.
+        :param bool recode_ancestral: If True, recode genotypes at sites where the
+            ancestral state is known such that the ancestral state is coded as 0,
+            as described above. Otherwise return genotypes in the input allele encoding.
+            Default: ``None`` treated as ``False``.
+        :return: An iterator over the variants in the sample data file.
+        :rtype: iter(:class:`Variant`)
+        """
+        if recode_ancestral is None:
+            recode_ancestral = False
+        all_genotypes = chunk_iterator(self.data["call_genotype"], indexes=sites)
+        assert MISSING_DATA < 0  # required for geno_map to remap MISSING_DATA
+        for genos, site in zip(all_genotypes, self.sites(ids=sites)):
+            # We have an extra ploidy dimension when coming from sgkit
+            genos = genos.reshape(self.num_samples)
+            aa = site.ancestral_allele
+            alleles = site.alleles
+            if aa != MISSING_DATA and aa > 0 and recode_ancestral:
+                # Need to recode this site
+                alleles = site.reorder_alleles()
+                # re-map the genotypes
+                geno_map = np.arange(len(alleles) - MISSING_DATA, dtype=genos.dtype)
+                geno_map[MISSING_DATA] = MISSING_DATA
+                geno_map[aa] = 0
+                geno_map[0:aa] += 1
+                genos = geno_map[genos]
+            yield Variant(site=site, alleles=alleles, genotypes=genos)
+
+    def _all_haplotypes(self, sites=None, recode_ancestral=None):
+        # We iterate over chunks vertically here, and it's not worth complicating
+        # the chunk iterator to handle this.
+        if recode_ancestral is None:
+            recode_ancestral = False
+        aa_index = self.sites_ancestral_allele[:]
+        # If ancestral allele is missing, keep the order unchanged (aa_index of zero)
+        aa_index[aa_index == MISSING_DATA] = 0
+        gt = self.data["call_genotype"]
+        chunk_size = gt.chunks[1]
+        for j in range(self.num_individuals):
+            if j % chunk_size == 0:
+                chunk = gt[:, j : j + chunk_size, :]
+            indiv_gt = chunk[:, j % chunk_size, :]
+            for k in range(self.ploidy):
+                a = indiv_gt[:, k].T
+                if recode_ancestral:
+                    # Remap the genotypes at all sites, depending on the aa_index
+                    a = np.where(
+                        a == aa_index,
+                        0,
+                        np.where(
+                            np.logical_and(a != MISSING_DATA, a < aa_index), a + 1, a
+                        ),
+                    )
+                yield (j * self.ploidy) + k, a if sites is None else a[sites]
+
+
 @attr.s(order=False, eq=False)
 class Ancestor:
     """