Inbound TCP and Full-Duplex gRPC Hit Cloudflare Workers
Cloudflare has introduced native support for inbound TCP sockets and bi-directional gRPC in its developer platform to support low-latency AI agent workloads.
On August 3, 2026, Cloudflare introduced inbound TCP connections and enhanced gRPC functionality to its developer platform, expanding the baseline networking capabilities of Cloudflare Workers and Cloudflare Containers. The updates, currently in private beta, provide the low-latency network primitives necessary for persistent AI voice interfaces and complex microservices.
If you build systems relying on bi-directional streams, you can now abandon complex Envoy proxy bridges and REST fallbacks. The release enables edge compute layers to accept and route raw persistent connections natively.
Native Inbound TCP via Spectrum
The runtime now exposes a connect(socket) handler. This handler allows a Worker to directly accept an inbound TCP socket provided by Cloudflare Spectrum, the platform’s ingress proxy for non-HTTP traffic.
Developers interact with the raw TCP socket within the Worker environment using the standard WritableStream and ReadableStream APIs. The platform also supports socket forwarding. Inbound sockets can be passed directly from a Worker to a Durable Object or a Cloudflare Container. This maintains the continuous connection state required when you build real-time voice agents or deploy streaming transcription services.
Full-Duplex gRPC Capabilities
Cloudflare Containers now support full-duplex, bi-directional gRPC. You can forward an inbound TCP socket from the edge Worker directly to a gRPC server running inside a container, holding the two-way communication channel open for continuous streaming tasks like voice-to-text-to-voice pipelines.
For developers writing standard Worker code rather than managing raw containers, the platform introduces an automatic protocol translation layer. Workers natively serve unary and server-streaming gRPC API types. Developers write their logic using gRPC-web, and the platform handles the underlying conversion to and from standard gRPC for all incoming and outgoing requests.
| Networking Feature | Previous Limitation | August 2026 Update |
|---|---|---|
| Inbound TCP | HTTP/HTTPS only | Native support via connect(socket) |
| gRPC Bi-directional | Required REST fallbacks | Full-duplex in Cloudflare Containers |
| gRPC Translation | Manual proxy required | Automatic gRPC-to-gRPC-Web conversion |
| Outbound TCP | Supported via VPC (June 2026) | Integrates with inbound for full loop |
Compute Orchestration for AI
This networking expansion aligns with Cloudflare’s broader “Agents Week” initiative, which focuses on delivering infrastructure primitives for autonomous systems. These workloads demand stateful, persistent environments rather than isolated, short-lived function execution.
The TCP update launched concurrently with @cloudflare/computer, an agent runtime designed to orchestrate state between lightweight isolates and full Linux containers based on the exact compute demands of the workload. To further simplify polyglot architecture, Workers RPC now operates natively across both Python and JavaScript environments.
By resolving the historical inability to handle HTTP/2 bidirectional streaming, Cloudflare positions its edge network as a direct alternative to traditional always-on PaaS providers for high-utilization multi-agent coordination patterns.
The inbound TCP and gRPC features require developers to request access via Cloudflare’s private beta signup form. If your architecture currently relies on proxy layers or REST polling to simulate persistent connections, you should evaluate your microservices to determine which routing layers can be simplified through direct edge socket termination.
Get Insanely Good at AI
The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.
Keep Reading
How to Cut CPU Costs with Cloudflare Workers Cache
You will learn how to configure Cloudflare Workers Cache to serve responses directly from entrypoints, handle invalidations, and partition cache keys.
Live Object Sharing Connects Python and JS Cloudflare Workers
Cloudflare's Workers RPC now allows Python and JavaScript runtimes to invoke remote methods and exchange live objects without manual serialization.
TML-Interaction-Small Achieves 0.40s Full-Duplex Latency
Thinking Machines Lab has released a research preview of TML-Interaction-Small, a 276-billion-parameter Mixture-of-Experts model for full-duplex conversation.
Cloudflare Ships Panic and Abort Recovery for Rust Workers
Cloudflare updated Rust Workers to support WebAssembly exception handling, preventing isolated panics from crashing entire serverless instances.
Build Real-Time Voice Agents with Cloudflare Agents SDK
Learn how to integrate low-latency voice interactions into your AI agents using Cloudflare's new @cloudflare/voice package and Durable Objects.