When the connection arrives, the handler gets an object with a stream pair on it. Cloudflare's own sample takes a writer off `socket.writable`, encodes `Hello, world!\n`, writes it and closes [6]. That is the whole entry-point surface. The interesting decisions come after: keep the socket in the Worker, hand it to a different Worker, or pass it to a Durable Object [4]. From the Durable Object, `getTcpPort()` delivers it into a Container, and that is where the full-duplex path terminates [5].
That is three deployment units for one bidirectional connection [22]. Which means the code speaking your protocol end to end lives in a Container process, with the Worker in front of it as ingress and routing. The demos make that legible: a Go gRPC echo server and a Python `socketserver`, both running unmodified [7]. Unmodified means an ordinary socket-listening process, which is a compliment to the plumbing rather than a claim about the isolate. Cloudflare's phrase for the result is full-duplex communication between client and server running any program, in any language, for any TCP-based protocol [8], and that sentence is accurate about the Container leg.
The Worker leg stops short, and the reason sits in the transport rather than in the roadmap. Developer code speaks gRPC-web, and Cloudflare converts incoming gRPC to gRPC-web and the outgoing direction back again [10]. gRPC's streaming, cancellation, flow control and trailers all depend on HTTP/2 frames carrying stream IDs, and web platform APIs such as `fetch()` do not expose that level [11]. Browsers hit the same wall, which is the reason gRPC-web exists at all [12]. Cloudflare has been converting gRPC down to HTTP/1.1 inside its reverse proxy since 2020 precisely so WAF rules and Bot Management can inspect the messages [13]. That history is baked into the transport, so a flag alone cannot undo it.
So the sharpest reply to the announcement is the operational one. Sebastian Buzdugan asked whether Workers expose enough backpressure control for gRPC streams, and noted that streaming edge cases are what decide whether an implementation holds; the post does not address it [14]. For a translated stream to survive a slow consumer, the consumer's read rate has to propagate back through the stream objects to the sender. Until that behaviour is documented and measured, server-streaming out of a Worker is an untested number in someone else's deployment.
On the cheap path the adoption cost is small. Existing clients need no changes, so a Worker can front an existing gRPC backend or serve mobile apps built on grpc-swift-2 and grpc-kotlin [15], and the server side is a few lines with the open-source `@connectrpc/connect` package [16]. The larger prize is the eight years of workloads the HTTP-only restriction excluded: message brokers, database proxies, custom binary protocols, anything expecting a socket [19]. A Worker now sees the connection before deciding where it goes, which is the shape of an API gateway applied to raw bytes, and what the Worker may inspect or enforce from that position is not covered in the announcement [20]. The voice-AI framing of Agents Week [21] describes one workload on this primitive, and a narrower one than most platform teams are likely to reach for first.
