diff --git a/api/DESIGN.md b/api/DESIGN.md
new file mode 100644
index 0000000000..0a2f36eb2b
--- /dev/null
+++ b/api/DESIGN.md
@@ -0,0 +1,71 @@
+# Design considerations
+
+The header files in this directory form the API to the WebRTC library
+that is intended for client applications' use.
+
+This API is designed to be used on top of a multithreaded runtime.
+
+The public API functions are designed to be called from a single thread*
+(the "client thread"), and can do internal dispatching to the thread
+where activity needs to happen. Those threads can be passed in by the
+client, typically as arguments to factory constructors, or they can be
+created by the library if factory constructors that don't take threads
+are used.
+
+Many of the functions are designed to be used in an asynchronous manner,
+where a function is called to initiate an activity, and a callback will
+be called when the activity is completed, or a handler function will
+be called on an observer object when interesting events happen.
+
+Note: Often, even functions that look like simple functions (such as
+information query functions) will need to jump between threads to perform
+their function - which means that things may happen on other threads
+between calls; writing "increment(x); increment(x)" is not a safe
+way to increment X by exactly two, since the increment function may have
+jumped to another thread that already had a queue of things to handle,
+causing large amounts of other activity to have intervened between
+the two calls.
+
+(*) The term "thread" is used here to denote any construct that guarantees
+sequential execution - other names for such constructs are task runners
+and sequenced task queues.
+
+# Client threads and callbacks
+
+At the moment, the API does not give any guarantee on which thread* the
+callbacks and events are called on. So it's best to write all callback
+and event handlers like this (pseudocode):
+<pre>
+void ObserverClass::Handler(event) {
+  if (!called_on_client_thread()) {
+    dispatch_to_client_thread(bind(handler(event)));
+    return;
+  }
+  // Process event, we're now on the right thread
+}
+</pre>
+In the future, the implementation may change to always call the callbacks
+and event handlers on the client thread.
+
+# Implementation considerations
+
+The C++ classes that are part of the public API are also used to derive
+classes that form part of the implementation.
+
+This should not directly concern users of the API, but may matter if one
+wants to look at how the WebRTC library is implemented, or for legacy code
+that directly accesses internal APIs.
+
+Many APIs are defined in terms of a "proxy object", which will do a blocking
+dispatch of the function to another thread, and an "implementation object"
+which will do the actual
+work, but can only be created, invoked and destroyed on its "home thread".
+
+Usually, the classes are named "xxxInterface" (in api/), "xxxProxy" and
+"xxx" (not in api/). WebRTC users should only need to depend on the files
+in api/. In many cases, the "xxxProxy" and "xxx" classes are subclasses
+of "xxxInterface", but this property is an implementation feature only,
+and should not be relied upon.
+
+The threading properties of these internal APIs are NOT documented in
+this note, and need to be understood by inspecting those classes.