diff --git a/tomograph/backends/zipkin/sender.py b/tomograph/backends/zipkin/sender.py
new file mode 100644
index 0000000..c24cdfb
--- /dev/null
+++ b/tomograph/backends/zipkin/sender.py
@@ -0,0 +1,105 @@
+import eventlet
+
+socket = eventlet.import_patched('socket')
+time = eventlet.import_patched('time')
+scribe = eventlet.import_patched('scribe.scribe')
+TTransport = eventlet.import_patched('thrift.transport.TTransport')
+TSocket = eventlet.import_patched('thrift.transport.TSocket')
+collections = eventlet.import_patched('collections')
+traceback = eventlet.import_patched('traceback')
+threading = eventlet.import_patched('threading')
+
+from thrift.protocol import TBinaryProtocol
+
+class ScribeSender(object):
+    def __init__(self, host='127.0.0.1', port=1463,debug=False,
+                 target_write_size=1000, max_write_interval=1.0, 
+                 socket_timeout=5.0, max_queue_length=50000, must_yield=True):
+        self.dropped = 0
+        self._remote_host = host
+        self._remote_port = port
+        self._socket_timeout = socket_timeout
+        self._hostname = socket.gethostname()
+        self._max_queue_length = max_queue_length
+        self._max_write_interval = max_write_interval
+        self._target_write_size = target_write_size
+
+        self._debug = True
+        self._log_buffer = collections.deque()
+        self._last_write = 0
+        self._must_yield = must_yield
+        self._lock = threading.Lock()
+        eventlet.spawn_after(max_write_interval, self.flush)
+
+    def close(self):
+        self.flush()
+
+    def send(self, category, msg):
+        """
+        Send one record to scribe.
+        """
+        log_entry = scribe.LogEntry(category=category, message=msg)
+        self._log_buffer.append(log_entry)
+        self._dropMsgs()
+        
+        now = time.time()
+        if len(self._log_buffer) >= self._target_write_size or \
+                now - self._last_write > self._max_write_interval:
+            self._last_write = now
+            eventlet.spawn_n(self.flush)
+            # We can't do a "looping call" or other permanent thread
+            # kind of thing because openstack creates new hubs.  The
+            # spawn_after here (and in __init__) give us a simple way
+            # to ensure stuff doesn't sit in the buffer forever just
+            # because the app isn't logging stuff.
+            eventlet.spawn_after(self._max_write_interval, self.flush)
+        if self._must_yield:
+            # prevent the flushers from starving
+            eventlet.sleep()
+
+    def flush(self):
+        dropped = 0
+        try:
+            self._lock.acquire()
+            buf = []
+            while self._log_buffer:
+                buf.append(self._log_buffer.popleft())
+            if buf:
+                if self._debug:
+                    print "ScribeSender: flushing {0} msgs".format(len(buf))
+                try:
+                    client = self._getClient()
+                    result = client.Log(messages=buf)
+                    if result == scribe.ResultCode.TRY_LATER:
+                        dropped += len(buf)
+                except:
+                    if self._debug:
+                        print "ScribeSender: caught exception writing log message:"
+                        traceback.print_exc()
+                    dropped += len(buf)
+        finally:
+            self._lock.release()
+            self.dropped += dropped
+            if self._debug and dropped:
+                print "ScribeSender: dropped {0} messages for communication problem.".format(dropped)
+
+    def _dropMsgs(self):
+        dropped = 0
+        while len(self._log_buffer) > self._max_queue_length:
+            self._log_buffer.popleft()
+            dropped += 1
+        self.dropped += dropped
+        if self._debug and dropped:
+            print "ScribeSender: dropped {0} messages for queue length.".format(dropped)
+
+    def _getClient(self):
+        # We can't just keep a connection because the app might fork
+        # and we'll be left with the parent process's connection.
+        # More robust to just open one for each flush.
+        sock = TSocket.TSocket(host=self._remote_host, port=self._remote_port)
+        sock.setTimeout(self._socket_timeout * 1000)
+        transport = TTransport.TFramedTransport(sock)
+        transport.open()
+        protocol = TBinaryProtocol.TBinaryProtocolAccelerated(transport)
+        return scribe.Client(protocol)
+