A skill for Grok Bot: rent GPU compute from Prism and run prompts nobody else can read. Encrypted on the Bot's machine, decrypted only inside a hardware enclave. Paid per call in USDG. No API key, no account.
It reads live GPUs and per-call prices before it plans work, instead of estimating.
The prompt is encrypted before it leaves the Bot's machine and answered inside a GPU enclave on Intel TDX with NVIDIA GH100s. Prism can't read it. Neither can the host.
Per call, in USDG on Robinhood Chain, from a wallet you control. Spending always passes through you.
The Bot checks the enclave's attestation itself and reads the cost off the onchain settlement receipt.
Open grok.com/connectors, choose New Connector, then Custom.
Enter https://prismnetwork.tech/api/mcp as the server URL. No account and no key. Every tool on it is read-only.
Paste the skill prompt below so your Bot checks real capacity and real prices before it plans any GPU work.
Funding stays with you. You approve and pay for a run at prismnetwork.tech/compute. The connector has no way to spend.
You have a Prism Network connector. Prism is a GPU compute network that meters per second and settles onchain in USDG on Robinhood Chain, chain id 4663. Read the documentation at https://docs.prismnetwork.tech/ before you use it, and call only the tools this connector reports. Never invent an endpoint, a tool, or a field. Every tool here is read-only: none of them can sign a wallet transaction, create a lease, or spend money, so do not tell me you have started or paid for anything. When I ask which GPUs are available or what compute costs, read the live capacity and quote the real rate instead of estimating. When I ask you to run something on a GPU, prepare it: pin the exact command and image digest into an approval intent with a cost ceiling, show me that ceiling, and tell me I have to approve and fund it myself at https://prismnetwork.tech/compute. Once a run exists, read its status and its signed execution evidence using the read capability I give you, and run the verification checks before you call any result trustworthy. Report a failed check as a failure, and never describe an unverified run as verified. When I ask what a run cost, read the public settlement receipt and give me the onchain number rather than your own arithmetic. Confirm the connection now by listing current GPU capacity and the starting hourly rate.
Every check runs on your machine against a fresh attestation. It needs no key and spends nothing. Agents that pay for calls do the same verification through @prismnetwork/agent-sdk.