diff --git a/scheduler/.gitignore b/scheduler/.gitignore
new file mode 100644
index 000000000..a65b41774
--- /dev/null
+++ b/scheduler/.gitignore
@@ -0,0 +1 @@
+lib
diff --git a/scheduler/README.md b/scheduler/README.md
new file mode 100644
index 000000000..8489d7870
--- /dev/null
+++ b/scheduler/README.md
@@ -0,0 +1,6 @@
+To upload newer version:
+
+```
+pip install -r requirements.txt -t lib
+appcfg.py update .
+```
diff --git a/scheduler/app.yaml b/scheduler/app.yaml
new file mode 100644
index 000000000..8bc59f004
--- /dev/null
+++ b/scheduler/app.yaml
@@ -0,0 +1,13 @@
+application: positive-cocoa-90213
+version: 1
+runtime: python27
+api_version: 1
+threadsafe: true
+
+handlers:
+- url: .*
+  script: main.app
+
+libraries:
+- name: webapp2
+  version: latest
diff --git a/scheduler/appengine_config.py b/scheduler/appengine_config.py
new file mode 100644
index 000000000..f4489ff96
--- /dev/null
+++ b/scheduler/appengine_config.py
@@ -0,0 +1,3 @@
+from google.appengine.ext import vendor
+
+vendor.add('lib')
diff --git a/scheduler/cron.yaml b/scheduler/cron.yaml
new file mode 100644
index 000000000..652aed802
--- /dev/null
+++ b/scheduler/cron.yaml
@@ -0,0 +1,4 @@
+cron:
+- description: periodic gc
+  url: /tasks/gc
+  schedule: every 5 minutes
diff --git a/scheduler/main.py b/scheduler/main.py
new file mode 100644
index 000000000..ed0c78e31
--- /dev/null
+++ b/scheduler/main.py
@@ -0,0 +1,112 @@
+import collections
+import json
+import logging
+import operator
+import re
+
+import flask
+from oauth2client.client import GoogleCredentials
+from googleapiclient import discovery
+
+from google.appengine.api import urlfetch
+from google.appengine.ext import ndb
+
+app = flask.Flask('scheduler')
+app.debug = True
+
+# We use exponential moving average to record
+# test run times.  Higher alpha discounts historic
+# observations faster.
+alpha = 0.3
+
+PROJECT = 'positive-cocoa-90213'
+ZONE = 'us-central1-a'
+
+class Test(ndb.Model):
+  total_run_time = ndb.FloatProperty(default=0.) # Not total, but a EWMA
+  total_runs = ndb.IntegerProperty(default=0)
+
+class Schedule(ndb.Model):
+  shards = ndb.JsonProperty()
+
+@app.route('/record/<path:test_name>/<runtime>', methods=['POST'])
+@ndb.transactional
+def record(test_name, runtime):
+  test = Test.get_by_id(test_name)
+  if test is None:
+    test = Test(id=test_name)
+  test.total_run_time = (test.total_run_time * (1-alpha)) + (float(runtime) * alpha)
+  test.total_runs += 1
+  test.put()
+  return ('', 204)
+
+@app.route('/schedule/<test_run>/<int:shard_count>/<int:shard>', methods=['POST'])
+def schedule(test_run, shard_count, shard):
+  # read tests from body
+  test_names = flask.request.get_json(force=True)['tests']
+
+  # first see if we have a scedule already
+  schedule_id = "%s-%d" % (test_run, shard_count)
+  schedule = Schedule.get_by_id(schedule_id)
+  if schedule is not None:
+    return flask.json.jsonify(tests=schedule.shards[str(shard)])
+
+  # if not, do simple greedy algorithm
+  test_times = ndb.get_multi(ndb.Key(Test, test_name) for test_name in test_names)
+  def avg(test):
+    if test is not None:
+      return test.total_run_time
+    return 1
+  test_times = [(test_name, avg(test)) for test_name, test in zip(test_names, test_times)]
+  test_times_dict = dict(test_times)
+  test_times.sort(key=operator.itemgetter(1))
+
+  shards = {i: [] for i in xrange(shard_count)}
+  while test_times:
+    test_name, time = test_times.pop()
+
+    # find shortest shard and put it in that
+    s, _ = min(((i, sum(test_times_dict[t] for t in shards[i]))
+      for i in xrange(shard_count)), key=operator.itemgetter(1))
+
+    shards[s].append(test_name)
+
+  # atomically insert or retrieve existing schedule
+  schedule = Schedule.get_or_insert(schedule_id, shards=shards)
+  return flask.json.jsonify(tests=schedule.shards[str(shard)])
+
+NAME_RE = re.compile(r'^host(?P<index>\d+)-(?P<build>\d+)-(?P<shard>\d+)$')
+
+@app.route('/tasks/gc')
+def gc():
+  # Get list of running VMs, pick build id out of VM name
+  credentials = GoogleCredentials.get_application_default()
+  compute = discovery.build('compute', 'v1', credentials=credentials)
+  instances = compute.instances().list(project=PROJECT, zone=ZONE).execute()
+  host_by_build = collections.defaultdict(list)
+  for instance in instances['items']:
+    matches = NAME_RE.match(instance['name'])
+    if matches is None:
+      continue
+    host_by_build[int(matches.group('build'))].append(instance['name'])
+  logging.info("Running VMs by build: %r", host_by_build)
+
+  # Get list of builds, filter down to runnning builds
+  result = urlfetch.fetch('https://circleci.com/api/v1/project/weaveworks/weave',
+    headers={'Accept': 'application/json'})
+  assert result.status_code == 200
+  builds = json.loads(result.content)
+  running = {build['build_num'] for build in builds if build['status'] == 'running'}
+  logging.info("Runnings builds: %r", running)
+
+  # Stop VMs for builds that aren't running
+  stopped = []
+  for build, names in host_by_build.iteritems():
+    if build in running:
+      continue
+    for name in names:
+      stopped.append(name)
+      logging.info("Stopping VM %s", name)
+      compute.instances().delete(project=PROJECT, zone=ZONE, instance=name).execute()
+
+  return (flask.json.jsonify(running=list(running), stopped=stopped), 200)
diff --git a/scheduler/requirements.txt b/scheduler/requirements.txt
new file mode 100644
index 000000000..d4d47e6eb
--- /dev/null
+++ b/scheduler/requirements.txt
@@ -0,0 +1,2 @@
+flask
+google-api-python-client