- Get Started

The Big Picture

At the highest level, OpenFlows divides the world into two halves: a control plane where the system's brain lives, and a work plane where disposable agents do the actual coding. Everything else in the architecture serves that split.

Diagram of the OpenFlows system architecture: a control plane (Coder server, the Controller brain, the coordination store, and a database) supervising a work plane of disposable worker workspaces, with a coordination relay joining the two.
The control plane supervises the work plane; a coordination relay joins them.

The control plane

The control plane is where the trusted, long-lived pieces live:

The work plane

The work plane is where the actual labor happens — and it is deliberately disposable and untrusted.

Worker workspaces contain no LLM keys, no raw repository credentials, and have tightly restricted network access. The intelligence they need is served to them from the control plane.

What connects the two

The two planes are joined by:

The mental model

If the control plane is a headquarters and a worker workspace is a job site, then OpenFlows is the management that plans the work, sends a crew to the site, inspects the result from a distance, and approves the delivery — while the job site itself has no copying machine for secrets and is torn down when the work is complete. Settling a messy run is the job of the orchestration loop, which reads reality, repairs what it can, and rises to a human when it cannot.

Two systems, two jobs

Coder governs where agents run; OpenFlows governs how they coordinate. Coder supplies the machines, identities, and the gateway where model calls happen. OpenFlows turns a pile of workspaces into a cooperative team that plans, reviews, recovers, and merges.

This is the whole architecture. Every other page in this section zooms into one corner of this picture.