diff --git a/jawa/df.py b/jawa/df.py
new file mode 100644
index 0000000..b1cfa09
--- /dev/null
+++ b/jawa/df.py
@@ -0,0 +1,273 @@
+# -*- coding: utf-8 -*-
+import itertools
+from functools import partial
+from struct import unpack
+from collections import namedtuple
+
+from jawa.cf import ClassFile
+from jawa.util.stream import DexStreamReader
+from jawa.util.utf import decode_modified_utf8
+from jawa.util.descriptor import method_descriptor
+"""
+Dalvik Executable support.
+
+The :mod:`jawa.df` module provides tools for working with Android DEX
+(``.dex``) files.
+"""
+
+
+FieldItem = namedtuple('FieldItem', 'origin type_ name')
+MethodItem = namedtuple('MethodItem', 'origin type_ name')
+ClassItem = namedtuple('ClassItem', [
+    'origin',
+    'access_flags',
+    'superclass',
+    'interaces',
+    'source_file',
+    'annotations_off',
+    'class_data_off',
+    'static_values_off'
+])
+
+
+class DexFile(object):
+    SUPPORTED_VERSIONS = ('035', '038')
+    ENDIAN_CONSTANT = 0x12345678
+    REVERSE_ENDIAN_CONSTANT = 0x78563412
+    NO_INDEX = 0xffffffff
+
+    def __init__(self, fio):
+        """
+        Implements the Android DEX format.
+
+        .. note::
+
+            Currently support is read-only. DexFile support is a work in
+            progress.
+
+        .. note::
+
+            All processing is done in-memory. Loading exceptionally large
+            DEX files may use large amounts of memory and time. Effort was
+            made to make usage simple over efficient.
+
+        :param fio: any file-like object providing ``.read()``.
+        """
+        # Only set when the DEX file was loaded from disk, in which case it's
+        # the adler32 checksum of the file - magic - sum.
+        self._checksum = None
+        self._signature = None
+        self._version = '038'
+        self._little_endian = True
+
+        self._link_size = 0
+        self._link_off = 0
+        self._map_off = None
+
+        self._string_table = []
+        self._type_table = []
+
+        self._from_io(fio)
+
+    def _from_io(self, fio):
+        """
+        Loads an existing DEX file from any file-like object supporting
+        seek(), read(), and tell().
+        """
+        read = fio.read
+
+        magic, version, null = unpack('4s3sc', read(8))
+        if magic != 'dex\n' or null != '\x00':
+            raise ValueError('invalid magic number')
+
+        if version not in self.SUPPORTED_VERSIONS:
+            raise ValueError('unsupported dexfile version')
+
+        self._version = version
+
+        # Before we read anything else we need to find the endianess
+        # flag or we may read things incorrectly.
+        fio.seek(0x28)
+        self._little_endian = unpack('<L', read(4))[0] == \
+            self.ENDIAN_CONSTANT
+        # ... and back up to the top.
+        fio.seek(8)
+
+        dsr = DexStreamReader(fio, self._little_endian)
+        self._checksum = dsr.uint()
+        self._signature = dsr.unpack('20B')
+        # Discard the file size, isn't useful.
+        dsr.uint()
+        # Discrad the header size, although this might actually be useful in
+        # the future to skip over unknown sections. Will future version headers
+        # always append? Who knows!
+        dsr.uint()
+        # We've already read the endian tag, discard it.
+        dsr.uint()
+
+        self._link_size = dsr.uint()
+        self._link_off = dsr.uint()
+        self._map_off = dsr.uint()
+
+        self._string_table = self._read_offset_list(
+            dsr,
+            self._read_string_item
+        )
+        self._type_table = self._read_offset_list(
+            dsr,
+            lambda dsr: dsr.uint()
+        )
+        proto_table = self._read_offset_list(
+            dsr,
+            self._read_proto_item
+        )
+        field_table = self._read_offset_list(
+            dsr,
+            self._read_field_item
+        )
+        method_table = self._read_offset_list(
+            dsr,
+            partial(self._read_method_item, proto_table=proto_table),
+        )
+        class_table = self._read_offset_list(
+            dsr,
+            self._read_class_item
+        )
+
+        for class_def in class_table:
+            pass
+
+    def _read_offset_list(self, dsr, item_callback):
+        """
+        Most of the header tables follow the same count + offset + list
+        format. This helper can be used to save some typing
+        """
+        count = dsr.uint()
+        offset = dsr.uint()
+
+        if offset == 0:
+            return
+
+        items = []
+        with dsr.jump(offset):
+            for i in itertools.repeat(None, count):
+                items.append(item_callback(dsr))
+
+        return items
+
+    def _read_string_item(self, dsr):
+        with dsr.jump(dsr.uint()):
+            # For some idiotic reason this is the length of the
+            # *decoded* string, with the encoded string being null
+            # terminated so read it in byte-by-byte until we find the
+            # end. We just discard the useless length.
+            dsr.uleb128()
+
+            c = bytearray()
+            while True:
+                v = dsr.ubyte()
+                if v == 0:
+                    break
+                c.append(v)
+
+            # Strings are stored as in the JVM using MUTF-8.
+            return decode_modified_utf8(c)
+
+    def _read_proto_item(self, dsr):
+        """
+        Reads the method prototype table.
+        """
+        # Discard the "shorty" descriptor, which we have no need of.
+        dsr.uint()
+        returns = self.string_table[self.type_table[dsr.uint()]]
+
+        parameters_offset = dsr.uint()
+        if parameters_offset != 0:
+            # If non-zero this prototype has a list of parameters
+            # at the offset.
+
+            with dsr.jump(parameters_offset):
+                parameters = [
+                    self.string_table[
+                        self.type_table[idx]
+                    ]
+                    for idx in self._read_type_list(dsr)
+                ]
+        else:
+            parameters = []
+
+        return method_descriptor(
+            # Create a "fake" JVM-style long form descriptor so we
+            # can re-use method_descriptor.
+            u'({0}){1}'.format(u''.join(parameters), returns)
+        )
+
+    def _read_type_list(self, dsr):
+        """
+        Reads a list of types, used in the prototype and class definition
+        tables.
+        """
+        return [dsr.ushort() for _ in itertools.repeat(None, dsr.uint())]
+
+    def _read_field_item(self, dsr):
+        return FieldItem(
+            self.string_table[self.type_table[dsr.ushort()]],
+            self.string_table[self.type_table[dsr.ushort()]],
+            self.string_table[dsr.uint()]
+        )
+
+    def _read_method_item(self, dsr, proto_table):
+        return MethodItem(
+            self.string_table[self.type_table[dsr.ushort()]],
+            proto_table[dsr.ushort()],
+            self.string_table[dsr.uint()]
+        )
+
+    def _read_class_item(self, dsr):
+        return ClassItem(
+            dsr.uint(),
+            dsr.uint(),
+            dsr.uint(),
+            dsr.uint(),
+            dsr.uint(),
+            dsr.uint(),
+            dsr.uint(),
+            dsr.uint()
+        )
+
+    @property
+    def checksum(self):
+        return self._checksum
+
+    @property
+    def version(self):
+        return self._version
+
+    @property
+    def little_endian(self):
+        return self._little_endian
+
+    @property
+    def link_off(self):
+        return self._link_off
+
+    @property
+    def link_size(self):
+        return self._link_size
+
+    @property
+    def map_off(self):
+        return self._map_off
+
+    @property
+    def string_table(self):
+        return self._string_table
+
+    @property
+    def type_table(self):
+        return self._type_table
+
+    @property
+    def signature(self):
+        """SHA-1 hash of the file minus the magic, checksum, and itself."""
+        return self._signature
diff --git a/jawa/util/stream.py b/jawa/util/stream.py
index 492b19b..6e89c3c 100644
--- a/jawa/util/stream.py
+++ b/jawa/util/stream.py
@@ -1,5 +1,25 @@
 # -*- coding: utf-8 -*-
-from struct import unpack_from, calcsize
+import types
+import contextlib
+from struct import unpack_from, calcsize, unpack
+
+
+class StreamWithStack(object):
+    def __init__(self):
+        self.position_stack = []
+
+    def push_pos(self):
+        self.position_stack.append(self.fio.tell())
+
+    def pop_pos(self):
+        self.fio.seek(self.position_stack.pop())
+
+    @contextlib.contextmanager
+    def jump(self, offset, from_what=0):
+        self.push_pos()
+        self.fio.seek(offset, from_what)
+        yield
+        self.pop_pos()
 
 
 class BufferStreamReader(object):
@@ -42,3 +62,69 @@ def read(self, length=None):
         r = self.buff[self.pos:self.pos+length]
         self.pos += length
         return r
+
+
+DEX_UNPACK_TEMPLATE = '''\
+def {name}(self, what=None):
+    return unpack('{endian}{fmt}', self.fio.read({size}))[0]
+'''
+
+
+class DexStreamReader(StreamWithStack):
+    """
+    A utility for reading IO streams using the same naming convention as the
+    official DEX language specs, along with efficiently supporting big & little
+    endian parsing.
+
+    Rather than constantly check to see if big endian parsing is flagged we
+    check once when created and then generate methods with the appropriate
+    parsing in them.
+    """
+    def __init__(self, fio, little_endian=True):
+        super(DexStreamReader, self).__init__()
+        self.fio = fio
+        self.endian = '<' if little_endian else '>'
+
+        # Since these methods are used so frequently we're doing our
+        # best to reduce per-call overhead by compiling little/big
+        # endian versions on the fly only once.
+        fmts = [
+            ('byte', 'b', 1),
+            ('ubyte', 'B', 1),
+            ('short', 'h', 2),
+            ('ushort', 'H', 2),
+            ('int', 'i', 4),
+            ('uint', 'I', 4),
+            ('long', 'q', 8),
+            ('ulong', 'Q', 8)
+        ]
+
+        for method_name, fmt, size in fmts:
+            definition = DEX_UNPACK_TEMPLATE.format(
+                name=method_name,
+                endian=self.endian,
+                fmt=fmt,
+                size=size
+            )
+
+            namespace = dict(unpack=unpack)
+            exec definition in namespace
+            setattr(self, method_name, types.MethodType(
+                namespace[method_name],
+                self
+            ))
+
+    def uleb128(self, why=None):
+        r = 0
+        v = 0
+
+        while True:
+            v = self.byte()
+            r = (r << 7) | (v & 0x7F)
+            if v > 0:
+                break
+
+        return r
+
+    def unpack(self, fmt, why=None):
+        return unpack(fmt, self.fio.read(calcsize(fmt)))