Architecture

OpenOrange is powered by Rivet, a scalable control plane that runs your Cloudflare workloads. Your code ships as standard Docker containers, so it runs anywhere Docker runs.

Any cloud AWS · Google Cloud · Azure · your servers

• Docker container

  Workers + Durable Objects

  on Node.js

• Docker container

  Workers + Durable Objects

  on Node.js

• Docker container

  Workers + Durable Objects

  on Node.js

Each holds your Workers, Durable Objects and the rest of your Cloudflare code, on Node.js.

connect out

Rivet control plane

Places each Actor on one of your containers, routes requests to it, and scales the load across them.

Open source · self-hosted, Rivet Cloud or BYOC

backed by

S3

Cheap, durable and scalable

• Your code, built into an image by orange deploy
• One Rivet Actor, such as a Durable Object
• Rivet
• Storage

Powered by Rivet, backed by S3

Rivet is an open-source control plane. Each Durable Object runs as a Rivet Actor: Rivet finds it, wakes it on one of your containers, and routes each request to it.

Rivet keeps state in S3, which is cheap, durable and scalable. Scaling out is a matter of adding machines. See Rivet Actors.

Standard Docker containers, in any cloud

orange deploy builds your Worker and the OpenOrange runtime into a standard Docker image, so it runs anywhere Docker runs.

Each container holds your Workers, Durable Objects, Workflows and the rest of your Cloudflare code. The containers connect out to Rivet, which spreads the load across them. To add capacity, add containers.

Self-hosted is this architecture as drawn: you run the containers and Rivet. Rivet Cloud and BYOC run the same setup, operated by Rivet.

Deploy to	Runs on
Self-hosted	Any cloud, Hetzner or on-prem
Rivet Cloud	A dedicated cluster on AWS or Google Cloud, in the region you choose
BYOC	Your AWS, Google Cloud or Azure account

See Deploy and self-host.

Each binding maps to the best-fit service

Your bindings and your code stay the same. Stateful bindings run on Rivet Actors. The rest plug into services you already use.

Each binding’s page has how it runs and what differs: see the bindings overview.

Real Node.js, not the Workers runtime

Your Worker runs in a Node.js process, so you’re not limited to the part of Node.js that the Workers runtime supports. Node.js APIs are always on, without nodejs_compat, and the machines you choose set the memory and CPU time.

That opens up the Node.js ecosystem: npm packages that need Node.js APIs the Workers runtime doesn’t have. See OpenOrange vs workerd.

AI context Full service map and runtime limits

Written for LLMs and coding agents: every edge case, measurement, and known deviation. You do not need it to use the feature.

Full service map

Every Cloudflare service, with its status and what differs.

Cloudflare	OpenOrange	Status	Notes
Workers	OpenOrange runtime	Drop-in	The Workers API surface (fetch handler, env bindings, waitUntil, Cache) on Node.js. Your code and wrangler config load unchanged.
Durable Objects	Rivet Actors	Partial	One Rivet Actor per object, with SQLite storage and alarms. Same class, same bindings.
Workers KV	Actor-backed KV	Drop-in	The KV binding API, one Actor per namespace. Strongly consistent instead of eventually.
Queues	Actor-backed queues	Drop-in	Producers and consumers keep the Queues API, with batching, retries and dead-letter queues. One batch runs at a time per queue.
Workflows	Cloudflare's own Workflows engine	Partial	The engine that wrangler dev runs, copied from @cloudflare/workflows-shared, with one Rivet Actor per instance. Steps, sleeps, events and retries work. instance.subscribe() is not available.
Cron Triggers	Actor schedules	Drop-in	Same cron syntax in your config. Crons fire in orange dev too.
D1	SQLite on Rivet	Drop-in	The D1 binding API on one SQLite database in a Rivet Actor.
R2	A Rivet Actor, or your S3-compatible store Amazon S3 Google Cloud Storage Cloudflare R2 Backblaze B2 Locally: Actor storage, MinIO	Pluggable	A bucket is an Actor by default. Point it at any S3-compatible store instead.
Workers AI	An OpenAI-compatible provider OpenRouter OpenAI Groq Together AI Fireworks AI Locally: Ollama	Pluggable	The AI binding calls any OpenAI-compatible endpoint you already pay for. Text and embedding models.
Sandboxes	Your sandbox provider E2B Daytona Freestyle Locally: agentOS VM, Local directory	Pluggable	The Sandbox SDK runs on a hosted provider, an agentOS VM or a local directory. Other Container images are not run.
Hyperdrive	Direct connection Neon Supabase Amazon RDS PlanetScale Locally: Postgres, MySQL	Drop-in	Works unchanged. You are already in the VPC, so env.HYPERDRIVE passes the connection string straight through.
Browser Run	Your browser Browserbase Browserless Steel Locally: Headless Chrome	Pluggable	@cloudflare/puppeteer drives your Chrome or a hosted browser over the DevTools Protocol.
Email	Your SMTP server Resend Amazon SES Postmark SendGrid Mailgun Locally: Mailpit	Pluggable	send_email sends through your SMTP server. Inbound mail reaches email() from your provider's webhook.
Vectorize, Images, Pipelines	Bring your own	Not supported	The bindings load and every call throws. Call a replacement such as pgvector with fetch() today.

Drop-in: the binding API works unchanged. Pluggable: the binding works unchanged and you choose the backend. Partial: it runs, but some documented APIs throw or act differently. Not supported: calls throw.

Runtime limits, compared

OpenOrange next to Cloudflare’s Workers Paid plan.

Limit	OpenOrange	Cloudflare (Workers Paid)
Memory per Worker	The Node.js heap of the runtime process	128 MB
CPU time per request	No limit	30 s default, 5 min max
Subrequests per request	OS limits	10,000
Alarm wall time	About 24.8 days, the longest Node.js timer	15 min
Node.js APIs	Node.js modules, with workerd-shaped process and fs	nodejs_compat subset
Native npm modules	Not tested	No

Cloudflare figures checked 2026-10-08.