Serveur MCP VoltageGPU

voltagegpu-cli

Deploy and manage VoltageGPU machines from your terminal or from Python:

• Confidential VMs: Intel TDX trust domains with an NVIDIA GPU, only your SSH keys on the machine, and an attestation you generate and verify yourself.
• Standard pods: whole GPU machines without an enclave, for workloads that are not sensitive.

Billing is per second. The first hour is charged when you deploy; when you stop a standard pod, the unused part of that hour goes back to your balance.

pip install voltagegpu-cli

Python 3.10 or newer. The package installs the voltagegpu command (volt is an alias kept from earlier releases).

1. Log in

Create an API key at https://voltagegpu.com/api-keys (it is shown once), then:

voltagegpu login                 # prompts for the key, hidden, checks it, saves it to ~/.volt/config.ini (mode 0600)
# or, for CI and scripts:
export VOLT_API_KEY=volt_...

Add the SSH key you will connect with (defaults to ~/.ssh/id_ed25519.pub):

voltagegpu ssh-keys add
voltagegpu ssh-keys list

2. Pick a machine

voltagegpu machines                          # Confidential VM tiers, then standard machines, cheapest first
voltagegpu machines --standard --gpu A6000   # filter by GPU model
voltagegpu machines --standard --max-price 1 --json
voltagegpu machines --confidential
voltagegpu templates                         # images for standard pods (PyTorch by default)

3a. Deploy a standard pod

voltagegpu deploy standard --machine <machine_id>
voltagegpu deploy standard --cheapest --gpu 4090 --wait     # cheapest matching machine, wait for SSH

The CLI shows the price and asks before charging. In scripts, add --yes. If your account has several SSH keys, choose one with --ssh-key <key_id>. Other options: --template <template_id>, --jupyter, --ports 3, --auto-terminate 8 (hours), --name my-pod.

3b. Deploy a Confidential VM

voltagegpu deploy confidential --tier rtx6000b-small --wait
voltagegpu deploy confidential --tier h100-small --hardstop 4 --port 8000 --port 8443/tcp

Every SSH key on your account is bound to the VM when it is created; a VM cannot be given a key afterwards, so add yours first. The sudo password is printed once and is not kept anywhere. The command also prints how to generate the attestation inside the VM.

4. Connect, inspect, stop

voltagegpu pods list
voltagegpu pods wait <pod_id>                 # until SSH answers
voltagegpu ssh <pod_id>                       # interactive shell
voltagegpu ssh <pod_id> -- nvidia-smi -L      # run one command
voltagegpu ssh <pod_id> --print               # just print the ssh command
voltagegpu pods show <pod_id>
voltagegpu pods stop <pod_id>                 # releases it: billing stops, the disk is deleted
voltagegpu balance

Fresh pods present a new host key. In non-interactive scripts you can accept it on first use with -o StrictHostKeyChecking=accept-new.

5. Verify a Confidential VM

Inside the VM you produce an evidence bundle (the deploy output prints the exact commands). On your own machine:

pip install "voltagegpu-cli[verify]"          # installs voltage-verify
voltagegpu verify bundle.json --challenge <hex-you-issued> --gpus 1
voltagegpu verify bundle.json --challenge <hex-you-issued> --offline --accept-offline-tcb

voltagegpu verify runs voltage-verify, which checks the Intel TDX quote up to Intel's root and the NVIDIA GPU attestation tokens, without trusting VoltageGPU. Online, the TCB status and revocation are checked against Intel's collateral of today. --offline uses the Intel collateral inside the bundle: the verdict names that limit and the exit code is 3, unless --accept-offline-tcb (voltage-verify 0.3.0 or later). voltagegpu verify returns voltage-verify's own exit code.

Scripting

Every list command takes --json; deploys and stops take --yes.

# Deploy the cheapest standard machine, wait for SSH, run a job, release it.
POD=$(voltagegpu deploy standard --cheapest --max-price 1 --yes --json | jq -r .deploy.pod_id)
voltagegpu pods wait "$POD"
voltagegpu ssh "$POD" -o StrictHostKeyChecking=accept-new -- nvidia-smi
voltagegpu pods stop "$POD" --yes

Exit codes: 0 success, 1 API or runtime error, 2 usage error or missing confirmation.

vgpu: the short command for scripts and AI agents

Installed with the package, next to voltagegpu. The key comes from VOLT_API_KEY or VOLTAGEGPU_API_KEY (or voltagegpu login) and is never printed. Give an agent a limited key (scopes, its own machines only, a daily budget) from https://voltagegpu.com/api-keys, never your full key.

vgpu ls --json                                   # every offer, cheapest first
vgpu ls --kind vm --gpu H100                     # vm | cpu | standard | baremetal | sandbox
vgpu up rtx6000b-small --name train1 --ttl 2h -y --json
vgpu ps --json
vgpu exec train1 "nvidia-smi"                    # returns the remote exit code
vgpu exec train1 -d "python train.py"            # background: prints PID and log path
vgpu scp train1 /home/ubuntu/out.tgz ./out.tgz   # --up to copy the other way
vgpu attest train1                               # Confidential VM or Confidential CPU only
vgpu rm train1 -y
vgpu whoami                                      # scopes, budgets, spend, balance
vgpu mcp                                         # local MCP server (stdio) for agents

--ttl is mandatory and becomes the auto-stop of each product, rounded up, refused locally when out of bounds: Confidential VM and CPU up to 720h, standard pods 1h to 720h (whole hours), dedicated machines 1h to 72h (whole hours), sandboxes up to 7d. up asks for confirmation unless -y is given, and fails without -y when not run in a terminal. A deploy that gets no answer is never retried: run vgpu ps first.

exec, scp and attest use your local SSH key (-i, or VGPU_SSH_KEY). The first connection to a machine records its host key in ~/.volt/known_hosts.d/<id> (trust on first use); a changed key is refused. Names given at up are kept in ~/.volt/vgpu-names.json; ids work too.

Confidential VMs and CPU tiers bind the SSH keys already on the account. A standard GPU binds one key: the account's only key, or --ssh-key-id ID when the account has several (voltagegpu ssh-keys list shows the ids). Dedicated machines and sandboxes take key ids: by default every key of the account, but a key limited to its own machines cannot read them, so pass --ssh-key-id ID (repeatable). For such a key whoami and voltagegpu balance show the balance as hidden: only a full key reads it.

vgpu attest draws a challenge on your machine, runs the published proof script in the machine (two-proofs-in-the-vm.sh for a Confidential VM, snp-proof-in-the-vm.sh for a Confidential CPU tier, both embedded and pinned by SHA-256), copies the archive back and verifies it here: voltage-verify for Intel TDX + NVIDIA (pip install 'voltagegpu-cli[verify]'). --offline verifies without network, with the Intel collateral inside the bundle: the TCB and revocation are then not checked against Intel today, so the result stays VERIFIED but names that limit (limits in --json) and exits with 10, unless --accept-offline-tcb. For SEV-SNP: the report is requested at --vmpl (0 by default, sent as SNP_VMPL; if VMPL 0 is refused, the script says so and asks again at 1, and the result reports it), REPORT_DATA equals the challenge, the report policy forbids debug and its VMPL field is the level requested (the guest kernel uses VMPCK0, consistent with VMPL 0, no SVSM), then the AMD chain and the VCEK are fetched again from AMD into a new folder (snpguest 0.11.0 fetch ca / fetch vcek on your machine), the ARK must be one of the AMD roots pinned in the script, and snpguest verify certs / verify attestation run with those certificates, never the ones in the archive (they come from the VM). Without a local snpguest, or when the AMD key server cannot be reached, the result is INCOMPLETE, never VERIFIED. Standard pods, dedicated machines and sandboxes have no tenant attestation and are refused.

Exit code	Meaning
0	ok
1	error (including a deploy that got no answer)
2	usage: bad option, --ttl out of bounds, missing -y
3	API key missing, malformed or refused
4	API key scope missing (API_KEY_SCOPE_MISSING), or account resource refused to a key limited to its own machines (API_KEY_ACCOUNT_RESOURCE)
5	not found (machine or tier)
6	balance too low or API key budget exceeded (API_KEY_BUDGET_EXCEEDED, HTTP 402): stop, nothing was charged
7	unavailable: sold out, over --max-price, rate limited, offer closed
8	attestation failed
9	attestation incomplete, or a local tool is missing (ssh, scp, voltage-verify)
10	attestation verified with a limit: --offline, TCB and revocation not checked against Intel today (--accept-offline-tcb accepts it; local MCP server: ATTESTATION_LIMITED)

With --json, an error is also printed on stdout as {"error": {"code", "message", "exit_code", ...}}.

Local MCP server

vgpu mcp serves the tools ls, up, ps, exec, scp, rm, attest, inference and whoami over stdio. It runs on your machine because exec, scp and attest use your SSH private key, which never leaves it. Same guards as the command line: up requires ttl and max_hourly_price, rm requires confirm: true, a 402 comes back as API_KEY_BUDGET_EXCEEDED with "do not retry", the key is never part of an answer, and the sudo password of a Confidential VM is never handed to the agent. inference calls the OpenAI-compatible API with an inference key (vgpu_...) in VOLTAGEGPU_API_KEY; a vgpu_ value there is never sent to the pod API, which takes the volt_ key from VOLT_API_KEY.

claude mcp add voltagegpu -e VOLT_API_KEY=volt_... -- vgpu mcp    # Claude Code

# Codex: ~/.codex/config.toml
[mcp_servers.voltagegpu]
command = "vgpu"
args = ["mcp"]
env = { VOLT_API_KEY = "volt_..." }

Python SDK

from volt.sdk import VoltageGPUClient

with VoltageGPUClient() as client:            # VOLT_API_KEY or ~/.volt/config.ini
    machine = client.list_machines("standard")[0]          # cheapest first
    result = client.deploy_standard(machine.id, name="my-pod")
    pod = client.wait_until_reachable(result.pod_id)
    print(pod.ssh_command)
    print(client.release_pod(result.pod_id).prepaid_refund)

A complete example is in examples/deploy_standard_pod.py.

REST API

The CLI is a thin client of the public API, documented at https://docs.voltagegpu.com/api-reference. The same deploy with curl:

curl -X POST https://api.voltagegpu.com/api/volt/pods \
  -H "X-API-Key: $VOLT_API_KEY" -H "Content-Type: application/json" \
  -d '{"provider":"standard","name":"my-pod","machine_id":"<machine_id>"}'

curl https://api.voltagegpu.com/api/volt/pods -H "X-API-Key: $VOLT_API_KEY"

Configuration

Setting	Where
API key	--api-key of login, VOLT_API_KEY (or VOLTAGEGPU_API_KEY), then ~/.volt/config.ini
API base URL	VOLT_BASE_URL (https only, voltagegpu.com hosts unless VOLT_ALLOW_CUSTOM_BASE_URL=1; default https://api.voltagegpu.com/api)
Debug	VOLT_DEBUG=1 shows tracebacks

voltagegpu config shows where the key comes from, masked.

Security

• The key is sent only over HTTPS, in the X-API-Key header; redirects are never followed and certificates are always verified.
• The key is never printed. ~/.volt/config.ini is created readable by you only, and voltagegpu config warns if its permissions are too open.
• The SSH command returned by the API is parsed into user, host and port, only that rebuilt form is ever shown, and ssh is started without a shell (and never from the current directory), so a tampered answer cannot run a command on your machine. Server text is printed literally, without terminal escapes.
• ssh-keys add refuses private keys.

Report a vulnerability to contact@voltagegpu.com.

Upgrading from 1.2

• volt cc inventory / volt cc deploy: Confidential containers were retired on 2026-10-01. Use voltagegpu machines --confidential and voltagegpu deploy confidential --tier <tier>.
• volt pods stop now releases the pod and reports the refund.
• volt account balance is now voltagegpu balance.

Links

• Docs: https://docs.voltagegpu.com/cli
• API reference: https://docs.voltagegpu.com/api-reference
• API keys: https://voltagegpu.com/api-keys
• Changelog: CHANGELOG.md

MIT License.