Connect an AI assistant to Theta EdgeCloud

Theta EdgeCloud has a hosted MCP server. Connect
Claude, Cursor or any other MCP-capable assistant to it once, and you can run AI
models, manage GPU Nodes and check your spending by asking in plain language.

Server URL: https://controller.thetaedgecloud.com/mcp

What your assistant can do

Run AI models on demand

"Which audio models do you have?"
"Transcribe this recording."
"Generate an image of a mountain sunset with Flux."
"Is that long video job done yet?"

Rent and operate GPU Nodes

"Which GPUs are available right now, and what do they cost?"
"Spin up an RTX 4090 node with my SSH key, but not above 50 cents an hour."
"Show me its logs."
"Stop it for the night." / "Start it again." / "Delete it."

Keep an eye on spending

"How much credit is left?"
"What did this project spend in August?"
"What does storage cost per GB?"

Starting a node costs money and deleting one is permanent, so the assistant
always asks you before doing either. Prices and permissions are exactly the
ones you see in the dashboard.

Connect

Claude Code

claude mcp add --scope user --transport http theta-edgecloud https://controller.thetaedgecloud.com/mcp

Start claude, type /mcp, pick theta-edgecloud and choose Authenticate.
Your browser opens the Theta EdgeCloud sign-in page described below.

Claude Desktop, Cursor and other MCP clients

Add this to your client's MCP configuration:

{
  "mcpServers": {
    "theta-edgecloud": {
      "type": "http",
      "url": "https://controller.thetaedgecloud.com/mcp"
    }
  }
}

The client opens the sign-in page in your browser on first use. There is no
client ID or secret to copy.

Scripts and automation

If you would rather not sign in through a browser, send your API keys as
headers instead. Use the on-demand API key for models and the project API key
for GPU Nodes and billing. Either or both:

claude mcp add --scope user --transport http theta-edgecloud https://controller.thetaedgecloud.com/mcp \
  --header "Authorization: Bearer YOUR_ONDEMAND_API_KEY" \
  --header "x-api-key: YOUR_PROJECT_API_KEY"

Connections made this way are not tied to a project, so the assistant will ask
which project to use.

Signing in and permissions

The sign-in page asks for:

Project ID, the project this connection can work in. One project per connection.
Project API key, for GPU Nodes and billing. Create one in the dashboard under
Settings → Projects → Create API Key.
On-demand API key, for running models. Create one on the On-demand Model APIs page.
Access to grant, one checkbox each:

Checkbox	What it allows	Default
`inference`	Run models: list them, run inference, check results	on
`gpu:read`	See available GPUs, templates, your nodes, their logs and events	on
`gpu:write`	Create, start, stop and delete GPU Nodes	off
`billing:read`	Balance, usage, top-ups and pricing	on

Only enter the keys the access you tick needs. The assistant only ever sees
what you allowed: without gpu:write it has no way to create or delete a
node at all.

Your keys are stored encrypted and never shown again. The connection stays
signed in for up to 90 days. To disconnect, remove the server from your
assistant, or delete the API key in the dashboard.

Already using the on-demand MCP package?

@thetalabs/on-demand-api-mcp keeps working. The hosted server offers the same
model tools plus GPU Nodes and billing, needs nothing installed, and works from
web-based assistants too. To switch, replace the npx entry in your config
with the server URL above and sign in.