AI overview
What the AI section does today, who can see it, and how to turn it on for your Loomplane server.
The AI section runs open models on your own servers and clusters. You pick a model from Loomplane's registry, or paste any Hugging Face repo, choose a GPU or a server, and Loomplane adds the model to your project's manifest and plans it like any other workload. Services in the same environment link to it and get two variables, MODEL_URL and MODEL_API_KEY, which they use to call it.
Everything runs on infrastructure you connected. Loomplane's control plane never hosts a model, and your services' requests to one never pass through Loomplane.
Turn it on
The section is off until an operator turns it on for the server. Set this and restart Loomplane:
export LOOMPLANE_AI_ENABLED=trueAn AI group then shows in the sidebar for everyone with ai:read, which is every role from Viewer up. It holds two pages:
- Models, where you browse the registry, deploy a model and see what each deployed model costs. See Models and What models cost.
- Settings, which holds the workspace's Hugging Face token for gated models. See AI settings.
AI settings also covers what happens to running models if you turn the section off again.
Who can do what
| Permission | Held by | What it allows here |
|---|---|---|
ai:read | Viewer and up | See the AI pages, the registry, deployed models and their costs |
ai:edit | Developer and up | Deploy, move, retry and remove models in environments that aren't production, and set their budget and price |
ai:manage | Admin and up | All of that in production too, and the Hugging Face token |
Creating a DigitalOcean GPU server for a model is a separate thing. It needs organization_cloud:manage, which admins hold, because the server bills from the moment it exists. And a model deploy makes a plan, so someone with deployment:approve still approves it before anything runs.
What isn't here yet
A gateway that serves several models behind one address, and agents that use them, are planned for later. Until the gateway exists, a model runs in the one environment you deploy it to, and only services in that environment link to it. If staging and production both need a model, deploy it to each, and each copy holds its own GPU.
The MCP server under AI agents is a different thing: it lets a coding agent such as Claude Code drive Loomplane. It works whether or not the AI section is on.
Notifications and webhooks
Get deployment results and the agent's reports by email, and send signed webhooks to your own systems.
Models
Browse the model registry, deploy a model in one click on a GPU or a server you have, run it in more than one environment, link services to it, and move or remove it later.