OpenDray: one control plane for local and cloud models

I run models everywhere. Local ones on the DGX and the Mac Studios, frontier ones in the cloud. The tooling to drive them was fragmented and single-model, and none of it treated a fleet of models the way you treat a fleet of anything in an enterprise: shared state, central secrets, standards-based integration, and a plan for the day a node fails. So I built that platform. It is called OpenDray, and it is open source.

One control plane, many models. Local and cloud models run side by side under a single control plane, not one tool per model. A task can go to a local model on the DGX for cost and privacy, or to a frontier model in the cloud for reach, and the way you work does not change.

A round table, not a queue. The models do not just take turns. Several can work the same problem together, review each other, and converge, so an answer is not hostage to one model’s blind spot.

A shared data layer. Every model reads and writes the same vector memory and the same knowledge base. Context and work persist across sessions and across models, so what one model learns, the next one already knows.

Central secrets. Credentials live in a shared key and secret store, not scattered across a dozen configs. Models get scoped access, and rotating a key is one operation rather than a hunt.

Skills written once. Capabilities live in one skill and plugin repository that every model can use. You build an integration once and the whole fleet has it, instead of re-implementing the same thing per agent.

Where people already are. It is driven from Telegram and Slack as first-class channels, so you can hand work to the fleet, or approve an action, from a phone.

Standards, not lock-in. Tool integration is over MCP, the Model Context Protocol, so it plugs into the wider ecosystem instead of a private API.

Built to survive. Backups and high availability are in from the start, because a platform you actually depend on cannot be a single point of failure.

I built it the way I build enterprise systems: shared state, central secrets, standards-based integration, backups, no single point of failure. Then I released it for anyone to run. It was never meant to be a demo. It is the infrastructure I use every day.