diff --git a/parso/python/fstring.py b/parso/python/fstring.py
new file mode 100644
index 0000000..e1d09f9
--- /dev/null
+++ b/parso/python/fstring.py
@@ -0,0 +1,130 @@
+import re
+
+from parso.utils import PythonVersionInfo
+from parso.python.tokenize import Token
+from parso.python import token
+
+version36 = PythonVersionInfo(3, 6)
+
+
+class TokenNamespace:
+    LBRACE = token.LBRACE,
+    RBRACE = token.RBRACE
+    ENDMARKER = token.ENDMARKER
+    ERRORTOKEN = token.ERRORTOKEN
+    COLON = token.COLON
+    CONVERSION = 100
+    PYTHON_EXPR = 101
+    EXCLAMATION_MARK = 102
+
+    @classmethod
+    def generate_token_id(cls, string):
+        if string == '{':
+            return cls.LBRACE
+        elif string == '}':
+            return cls.RBRACE
+        elif string == '!':
+            return cls.EXCLAMATION_MARK
+        return getattr(cls, string)
+
+
+GRAMMAR = """
+fstring: expressions ENDMARKER
+expressions: expression*
+expression: '{' PYTHON_EXPR [ '!' CONVERSION ] [ ':' expressions ] '}'
+"""
+
+_prefix = r'((?:[^{}]+|\}\}|\{\{)*)'
+_expr = _prefix + r'(\{|\}|$)'
+_in_expr = r'[^{}\[\]:"\'!]*(.?)'
+# There's only one conversion character allowed. But the rules have to be
+# checked later anyway, so allow more here. This makes error recovery nicer.
+_conversion = r'([^={}:]+)(.?)'
+
+_compiled_expr = re.compile(_expr)
+_compiled_in_expr = re.compile(_in_expr)
+_compiled_conversion = re.compile(_conversion)
+
+
+def tokenize(code, start_pos=(1, 0)):
+    def tok(value, type=None, prefix=''):
+        if type is None:
+            type = TokenNamespace.generate_token_id(found)
+        return Token(type, value, (line, column), prefix)
+
+    code = ''
+    start = 0
+    while True:
+        match = _compiled_expr.match(code, start)
+        prefix = match.group(1)
+        found = match.group(2)
+        start = match.end()
+        if not found:
+            # We're at the end.
+            break
+
+        if found == '}':
+            yield tok(found, prefix=prefix)
+        else:
+            assert found == '{'
+            yield tok(found, prefix=prefix)
+
+            expression = ''
+            squared_count = 0
+            curly_count = 0
+            while True:
+                expr_match = _compiled_in_expr.match(code, start)
+                expression += expr_match.group(0)
+                found = expr_match.group(2)
+                start = expr_match.end()
+
+                if found == '{':
+                    curly_count += 1
+                    expression += found
+                elif found == '}' and curly_count > 0:
+                    curly_count -= 1
+                    expression += found
+                elif found == '[':
+                    squared_count += 1
+                    expression += found
+                elif found == ']':
+                    # Use a max function here, because the Python code might
+                    # just have syntax errors.
+                    squared_count = max(0, squared_count - 1)
+                    expression += found
+                elif found == ':' and (squared_count or curly_count):
+                    expression += found
+                elif found in ('"', "'"):
+                    expression += found
+                    search = found
+                    if len(code) > start + 1 and  \
+                            code[start] == found == code[start+1]:
+                        search *= 3
+                        start += 2
+
+                    index = code.find(search)
+                    if index == -1:
+                        index = len(code)
+                    expression += code[start:index]
+                    start = index
+                elif found == '!' and len(code) > start and code[start] == '=':
+                    # This is a python `!=` and not a conversion.
+                    expression += found
+                else:
+                    yield tok(expression, type=TokenNamespace.PYTHON_EXPR)
+                    if found:
+                        yield tok(found)
+                    break
+
+            if found == '!':
+                conversion_match = _compiled_conversion.match(code, start)
+                found = conversion_match.group(2)
+                start = conversion_match.end()
+                yield tok(conversion_match.group(1))
+                if found:
+                    yield tok(found)
+
+            # We don't need to handle everything after ':', because that is
+            # basically new tokens.
+
+    yield tok('', type=TokenNamespace.ENDMARKER, prefix=prefix)