VoltageGPU runs a remote MCP server at https://voltagegpu.com/api/mcp with six tools: get_verification_instructions (public, the attestation commands and evidence index), list_vm_tiers (live inventory and price), register_ssh_key, deploy_vm, list_pods (status, ssh_ready, ssh_command), and stop_pod. This endpoint is an ordinary API route, not itself sealed in Intel TDX; the Confidential VM it provisions is what runs inside Intel TDX, with memory encrypted by the processor and isolated from the hypervisor and other tenants.
No separate MCP fee. Reading tools are free. deploy_vm charges the tier's hourly rate, one hour prepaid, refunded per second on stop_pod.
One remote MCP server, six tools: deploy an Intel TDX VM with an NVIDIA GPU, register a key, poll until it is reachable, get the exact commands to generate and verify both attestation proofs, then release it. No dashboard, no human in the loop.
To be precise about what is and is not sealed: this MCP endpoint is an ordinary API route, like any other on this site. What it lets an agent provision, the Confidential VM itself, is what runs inside Intel TDX. See what is confidential and what is not below, we are not going to blur that line to sound more impressive.
Model Context Protocol is an open standard for connecting LLM clients and coding agents to external tools. An MCP server exposes tools (callable functions with typed inputs and outputs) that a client can list and call over a standard transport. The full spec lives at modelcontextprotocol.io.
Agents are starting to buy and provision infrastructure on their own, not just reason about it. This server is our answer to that: everything a customer can do through the dashboard to run and verify a Confidential VM is also a tool call.
Point any MCP client that speaks Streamable HTTP at https://voltagegpu.com/api/mcp with your VoltageGPU API key as a bearer token. For a stdio-only client, wrap the same URL with mcp-remote.
// MCP client config (Claude Code, Cursor, Claude.ai custom connector, ...)
{
"mcpServers": {
"voltagegpu": {
"url": "https://voltagegpu.com/api/mcp",
"headers": {
"Authorization": "Bearer vgpu_live_..."
}
}
}
}# Any MCP client that speaks Streamable HTTP works the same way.
# This is the raw JSON-RPC call list_vm_tiers makes, for anything that
# is not yet a full MCP client:
curl -s https://voltagegpu.com/api/mcp \
-H "Authorization: Bearer $VOLTAGEGPU_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"jsonrpc": "2.0", "id": 1, "method": "tools/call",
"params": { "name": "list_vm_tiers", "arguments": {} }
}'Get a key at voltagegpu.com/api-keys. get_verification_instructions needs no key at all; every other tool returns a plain-text error naming the missing key rather than a silent empty result.
get_verification_instructions
Returns the exact commands to generate and verify both attestation proofs inside a Confidential VM, and which GPU SKUs already have a published, reproducible proof.
list_vm_tiers
Live inventory: resource name, GPU model and count, price per hour, how many are free right now, whether that SKU has a published attestation.
register_ssh_key
Registers a public key on the account. Do this before deploy_vm: a Confidential VM binds its keys at creation and cannot be given one afterwards.
deploy_vm
Deploys a Confidential VM on the tier you name. One hour is charged upfront at that tier’s rate, refunded per second on release.
list_pods
Status of every pod and VM on the account, including ssh_ready and the ready-to-use ssh_command once the machine actually answers.
stop_pod
Releases a pod or VM by id. Whatever remains of the prepaid hour comes back per second.
This endpoint is a normal API route
https://voltagegpu.com/api/mcp runs as an ordinary serverless function, the same as every other route on this site. It is not sealed in Intel TDX, and we are not going to claim otherwise to sound more impressive.
The VM it deploys is what runs inside Intel TDX
deploy_vm provisions an actual Confidential VM: hardware-encrypted memory, isolated from the hypervisor and other tenants. That guarantee lives on the VM, not on the tool call that requested it.
Both proofs are generated inside the VM, by you
get_verification_instructions returns the exact commands to produce an Intel TDX quote and an NVIDIA GPU attestation report on a nonce you choose, then verify both outside the VM. VoltageGPU is not in that trust chain.
Your API key is forwarded, never stored twice
Each tool call passes your bearer token straight through to the corresponding REST endpoint. This route holds no separate credential store and logs no token.
No separate MCP fee. get_verification_instructions, list_vm_tiers and list_pods cost nothing to call. deploy_vm charges exactly the tier’s hourly rate, one hour prepaid, refunded per second on stop_pod. Live rates are on the pricing page and returned live by list_vm_tiers.
EXPLORE FURTHER
Wire an agent to a real GPU in one config block
Generate an API key and call list_vm_tiers today.