diff --git a/docs/streaming.md b/docs/streaming.md
new file mode 100644
index 0000000..cc8b1e4
--- /dev/null
+++ b/docs/streaming.md
@@ -0,0 +1,131 @@
+# Streaming
+
+Connect supports several types of streaming RPCs. Streaming is exciting — it's fundamentally different from
+the web's typical request-response model, and in the right circumstances it can be very efficient. If you've
+been writing the same pagination or polling code for years, streaming may look like the answer to all your
+problems.
+
+Temper your enthusiasm. Streaming also comes with many drawbacks:
+
+- It requires excellent HTTP libraries. At the very least, the client and server must be able to stream HTTP/1.1
+  request and response bodies. For bidirectional streaming, both parties must support HTTP/2. Long-lived streams are
+  much more likely to encounter bugs and edge cases in HTTP/2 flow control.
+
+- It requires excellent proxies. Every proxy between the server and client — including those run by cloud providers —
+  must support HTTP/2.
+
+- It weakens the protections offered to your unary handlers, since streaming typically requires proxies to be
+  configured with much longer timeouts.
+
+- It requires complex tools. Streaming RPC protocols are much more involved than unary protocols, so cURL and your
+  browser's network inspector are useless.
+
+In general, streaming ties your application more closely to your networking infrastructure and makes your application
+inaccessible to less-sophisticated clients. You can minimize these downsides by keeping streams short-lived.
+
+All that said, `connect-python` fully supports client and server streaming. Bidirectional streaming is currently not
+supported for clients and requires an HTTP/2 ASGI server for servers.
+
+## Streaming variants
+
+In Python, streaming messages use standard `AsyncIterator` for async servers and clients, or `Iterator` for sync servers
+and clients.
+
+In _client streaming_, the client sends multiple messages. Once the server receives all the messages, it responds with
+a single message. In Protobuf schemas, client streaming methods look like this:
+
+```protobuf
+service GreetService {
+  rpc Greet(stream GreetRequest) returns (GreetResponse) {}
+}
+```
+
+In _server streaming_, the client sends a single message and the server responds with multiple messages. In Protobuf
+schemas, server streaming methods look like this:
+
+```protobuf
+service GreetService {
+  rpc Greet(GreetRequest) returns (stream GreetResponse) {}
+}
+```
+
+In _bidirectional streaming_ (often called bidi), the client and server may both send multiple messages. Often, the
+exchange is structured like a conversation: the client sends a message, the server responds, the client sends another
+message, and so on. Keep in mind that this always requires end-to-end HTTP/2 support!
+
+## HTTP representation
+
+Streaming responses always have an HTTP status of 200 OK. This may seem unusual, but it's unavoidable: the server may
+encounter an error after sending a few messages, when the HTTP status has already been sent to the client. Rather than
+relying on the HTTP status, streaming handlers encode any errors in HTTP trailers or at the end of the response body.
+
+The body of streaming requests and responses envelopes your schema-defined messages with a few bytes of
+protocol-specific binary framing data. Because of the interspersed framing data, the payloads are no longer valid
+Protobuf or JSON: instead, they use protocol-specific Content-Types like `application/connect+proto`.
+
+## An example
+
+Let's start by amending the `GreetService` we defined in [Getting Started](./getting-started.md) to make the Greet method use
+client streaming:
+
+```protobuf
+syntax = "proto3";
+
+package greet.v1;
+
+message GreetRequest {
+  string name = 1;
+}
+
+message GreetResponse {
+  string greeting = 1;
+}
+
+service GreetService {
+  rpc Greet(stream GreetRequest) returns (GreetResponse) {}
+}
+```
+
+After running `buf generate` to update our generated code, we can amend our service implementation in
+`server.py`:
+
+=== "ASGI"
+
+    ```python
+    from greet.v1.greet_connect import GreetService, GreetServiceASGIApplication
+    from greet.v1.greet_pb2 import GreetResponse
+
+    class Greeter(GreetService):
+        async def greet(self, request, ctx):
+            print("Request headers: ", ctx.request_headers())
+            greeting = ""
+            async for message in request:
+                greeting += f"Hello, {message.name}!\n"
+            response = GreetResponse(greeting=greeting)
+            ctx.response_headers()["greet-version"] = "v1"
+            return response
+
+    app = GreetServiceASGIApplication(Greeter())
+    ```
+
+=== "WSGI"
+
+    ```python
+    from greet.v1.greet_connect import GreetServiceSync, GreetServiceWSGIApplication
+    from greet.v1.greet_pb2 import GreetResponse
+
+    class Greeter(GreetServiceSync):
+        def greet(self, request, ctx):
+            print("Request headers: ", ctx.request_headers())
+            greeting = ""
+            for message in request:
+                greeting += f"Hello, {message.name}!\n"
+            response = GreetResponse(greeting=f"Hello, {request.name}!")
+            ctx.response_headers()["greet-version"] = "v1"
+            return response
+
+    app = GreetServiceWSGIApplication(Greeter())
+    ```
+
+That's it - metadata interceptors such as our [simple authentication interceptor](./interceptors.md#metadata-interceptors)
+can be used as-is with no other changes.