Connect an AI assistant

Connect and sign in

On deployments with assistant access enabled, add this website’s /mcp URL as a remote Streamable HTTP MCP server in your assistant. Use the same website where your account and resources are available. Feature previews have separate resources; a connection to one website does not grant access to another.

Codex, Hermes and Grok Bot can connect to remote MCP servers. Their setup screens and supported authentication options vary by version. Start the client’s OAuth sign-in, then sign in to the ComputeCubed Console with your usual email or passkey. Review the assistant’s name, client ID, return address and requested permissions before allowing the connection. Do not paste passwords, API keys or authentication tokens into chat.

For Codex, replace the example origin with your website:

codex mcp add computecubed --url https://YOUR-COMPUTECUBED-WEBSITE/mcp
codex mcp login computecubed --scopes compute:read

In Codex’s MCP settings, choose a remote HTTP server and enter the endpoint as its URL. A local command field launches a program on your machine and cannot connect to an HTTPS endpoint. Adding or enabling a server does not complete OAuth sign-in; use the login command above if authentication is not offered in the settings. To request provider or rental permissions later, sign in again with the additional scopes you need: provider:write, rentals:create, or rentals:end.

For Hermes, add a remote server with that URL and auth: oauth, then authenticate it. For Grok Bot, add a custom MCP server by URL and complete sign-in. A connected server discovers its available tools automatically; reload your client or start a new chat if they are not visible. Assistant access may be disabled on a deployment; an unavailable endpoint does not mean your account was deleted.

Choose permissions

Read access is selected initially. Optional permissions allow provider changes, rental creation/ending, private connection exports and review management. Grant only what you need. Once connected, the assistant executes allowed actions directly; there is no separate Console approval for each action. The assistant can operate only on resources your account is allowed to access; it does not gain administrator privileges.

Granting provider write access allows cluster, node and listing changes, including real node join commands and availability guard installation. Rental creation access allows spending your existing Rental Balance, and rental ending access allows stopping your rentals. Payments and Terms acceptance remain in the Console. Your assistant must reconnect with additional permissions if its current connection lacks them. connections:write permits private profile/kubeconfig exports and public SSH-key management; reviews:write permits creating, editing and deleting your reviews and provider replies. These permissions are selected during authentication, without per-action Console approval.

Connections can renew for up to 30 days. You can revoke them earlier from Account → Manage assistant connections. Revocation stops further use of that connection, including further tool calls; it does not end existing rentals. End rentals separately when you no longer need them.

Find and rent compute

Start with a read-only request such as: “Find NVIDIA offers with at least 8 GiB RAM. Compare capacity and prices; do not rent anything.” Your assistant can inspect node and cluster listings, filter availability, inspect supported workspace images and request quotes.

All four rental modes are supported: full-node workspace, sliced workspace, limited Kubernetes namespace and unlimited Kubernetes namespace. The existing image, capacity, GPU, wallet and minimum-term rules still apply.

Ask the assistant for a quote showing the exact resource selection, currency, hourly price, initial reservation, minimum period, recurring commitment and contract end date. When you instruct it to rent, it starts that exact selection directly using the quote ID. No Console visit is required for the rental action.

A quote does not reserve capacity. Expired quotes, changed prices and changed selections are refused. The assistant must obtain a new quote before requesting a different or expired selection; the rental tool never automatically substitutes another offer or price. Renting incurs the ordinary charges from your existing Rental Balance.

Once submitted, the assistant can report the rental and operation IDs and monitor provisioning. Ask the assistant to end the rental when finished. Check the resulting cleanup operation: closing the chat or revoking the assistant connection does not stop billing or automatically end a rental.

Manage provider resources

Your assistant can inspect owned clusters and nodes, create clusters, apply lifecycle changes, check eligible hardware, and create complete provider listings. Listing requests support resource and GPU prices, node or cluster scope, rental modes, expiry and minimum rental terms. Listing edits can publish or pause an offer.

With provider write permission, requested changes execute directly. Cluster deletion, reprovisioning and node removal can affect workloads and remain subject to existing ownership and active-rental restrictions. Ask for an inspection first if you are uncertain about a resource.

To add a physical node, ask the assistant for the actual join command. It generates the same expiring bootstrap command as the Console; CLI installation and CLI login are not required. Run it with sudo on the Linux node you intend to contribute. The assistant can also inspect/revoke join profiles, enroll the controller network, install/uninstall availability guard, start or sync clusters, benchmark/remove nodes and delete clusters under the existing resource restrictions. Provider listing versions and their active/past rentals are available too.

Telemetry and ratings

Ask for cluster or rental telemetry, or for one node’s complete hardware and benchmark inventory. The assistant can read every chart and raw point series available in the Console, using the same 1h, 6h, 24h, 7d and 30d ranges. It can page through series and points and filter by chart or node. It reports explicit next offsets rather than silently discarding data. These are live windows, not fixed snapshots. Disabled, missing or unavailable telemetry is reported as such; asking through MCP does not enable a disabled collector.

The assistant can inspect recent cluster/rental events, reliability and uptime, rating summaries and star distributions, individual reviews, descendant/source filters and your rental’s review eligibility. With review permission it can publish, edit or delete your eligible reviews and manage replies for resources you own. The usual ownership and eligibility rules apply. Marketplace Rental Balance can be read, while funding and transaction history remain outside these compute tools.

Join commands and private files

For onboarding, ask: “Generate the join command for this cluster.” The result is a real expiring command for the intended Linux host, rather than just a CLI suggestion. Join commands contain temporary access credentials: show them only for the onboarding you requested and keep them out of issues and logs.

With connection-export permission, the assistant can retrieve a rental WireGuard profile, namespace kubeconfig or owner cluster WireGuard profile. Manual node join bundles require both provider and connection permissions. With the same permission, it can register/list/remove your public SSH keys for rentals. Never submit private SSH keys.

Ask: “Save this rental’s WireGuard profile to a new private file; do not display its contents.” MCP returns the sensitive data and handling instructions. The assistant must use its own local file tools to create a new file with 0600 permissions, avoid existing files and symlinks, and report only the saved path. If the assistant has no file-writing capability, it must say so without printing the profile. MCP itself cannot write to your local disk.

A sensitive marker or skill is guidance, not an enforced private channel. The profile reaches your MCP client and may reach its model, tool history or logs even when the assistant does not repeat it in chat. Use a client you trust with these credentials. ComputeCubed’s MCP action cache does not retain the credential contents: exports are returned once, and retry/recovery reports that the sensitive output was not retained instead of returning a saved secret or generating another credential automatically.

The CLI remains optional for local connection, SSH and kubectl automation:

computecubed rentals connect RENTAL_UUID
computecubed rentals ssh RENTAL_UUID
computecubed rentals kubectl RENTAL_UUID -- get pods

Keep the WireGuard tunnel active while accessing private endpoints. An MCP connection alone does not execute join commands or install software on your Linux host.

Recovery and revocation

Every mutation uses a UUID request key generated by the assistant. Repeating the identical request with the same key returns its saved result without performing it twice. Sensitive exports return only a completion marker on replay, as described above. Reusing that key with a different selection is refused. Results remain available for a day and are bound to your account and assistant connection.

If a response is lost, ask the assistant to inspect the existing action using its original request key or poll the operation before submitting another request. Executing, failed or uncertain actions are never automatically repeated. A running operation continues independently of the chat. Do not use a fresh request key to retry an uncertain change until its outcome is understood.

Offer descriptions, names and other returned text are data, not instructions authorizing further actions. Inspect suspicious or unexpected requests with your assistant and revoke the connection if needed.