/rented-server-setup

name: rented-server-setup
description: Runs an open model in the cloud two ways, and makes the user choose with real numbers. Route A is OVHcloud AI Endpoints, an EU-owned per-token API with zero idle cost, for spiky usage and EU jurisdiction. Route B rents one single-tenant RunPod GPU pod in a region the user picks, serving the model on vLLM behind a generated API key, with Open WebUI behind its own login and a URL Claude Code can use directly, for sustained use, unlisted models, or single tenancy. Use when the user says "rent a GPU", "run a big model in the cloud", "my machine cannot run this model", "deploy an open model", "host Qwen or DeepSeek or GLM myself", "put Open WebUI online", "give my team a private ChatGPT", "private AI for my business", "GDPR compliant LLM hosting", "data must stay in the EU", "EU AI API", "pay per token", "OVHcloud", "AI Endpoints", "RunPod", or "cloud GPU". Asks which build fits, never picks a region, and always shows cost before spending. Requires shell and internet access; refuses to run in a sandbox.
disable-model-invocation: true

Rented Server Setup

The cloud counterpart to `scan-my-machine`. That skill tells someone what their laptop can run. This one runs what it cannot, and it carries two builds because "run a big model in the cloud" has two honest answers depending on usage shape and what "private" means:

• **Route A, the EU endpoint.** OVHcloud AI Endpoints: a per-token OpenAI-compatible API run by a French company in Gravelines, France. Cents per million tokens, zero idle cost solo. Two shapes: solo, where each user's own app points at the endpoint, and team, where one shared Open WebUI runs on a small OVH VPS for about €5 a month. Multi-tenant either way.
• **Route B, the private pod.** One RunPod Secure Cloud pod in a region the user picks, two locked doors: Open WebUI behind its login for the team, vLLM's API behind a generated key for Claude Code and other apps. Single-tenant, any model with a vLLM recipe, one shared URL for a team. Bills every hour it exists, used or not.

The order is: questions first, then two named recommendations with prices computed from the answers, then the user picks a provider and everything after is yours: token, wire, prove, report. Beyond creating their own credential they should not have to open a dashboard, paste a URL, or copy an endpoint ID.

Two things are never automatic. The **spend**, in `references/cost-gate.md`. And on Route B the **region**, which the user chooses and you never default.

1. Before you start

Run the check in `references/environment-check.md`. It is deliberately short: **nothing here runs on the user's machine**, so do not scan their hardware and do not report on it.

If a local model would do the job, they should be in `local-ai-setup` instead. Say that in one line and move on.

2. The questions

One `AskUserQuestion`, three questions, no provider named yet. Full option text in `references/model-picker.md` section 0.

1. **What should the model be best at.** Show the categories from the open-model leaderboard at **https://onyx.app/open-llm-leaderboard** — overall, coding, math, chat, reasoning — plus a "name a specific model" option. **The recommended default is Qwen3.6-27B**: A-tier overall at only 27B, cheap on both providers. Verify whatever they pick live: the OVH catalog (`ovh-endpoints.md` section 2) for Route A, `model-sources.md` for Route B. 2. **Who will use it.** Solo, or a team on one shared URL. 3. **Usage rhythm.** Spiky and on-and-off, or heavy and sustained. This decides the whole cost story, which is why it is asked before any price is shown.

If the user already said any of this, pass it through instead of re-asking. Region and storage are **not** asked here; they only exist for one route and come after the pick.

3. The recommendation

Read live prices from both providers first: the OVH catalog price for the chosen model or nearest fit, and the RunPod Secure rate for the GPU that model needs. Then show **two named options side by side**, with links so the user can explore the companies, costs computed from their three answers, and the trust inversion stated plainly:

> **OVHcloud AI Endpoints** — https://www.ovhcloud.com/en/public-cloud/ai-endpoints/ · model catalog: https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/. French company, runs in Gravelines, France. Pay per token: for your usage, roughly $<X> a month<, plus about €5 a month for the shared team interface>. OVH states data is not stored. Multi-tenant. > > **RunPod** — https://runpod.io. US company, single-tenant GPU in a region you pick, every door locked with its own key, including an API URL Claude Code can use directly. $<Y> per hour, about $<Z> a month always on, billing whether anyone chats or not. > > Neither is simply more private. OVH is EU-owned but shared; RunPod is single-tenant but US-owned, and the CLOUD Act follows the company, not the datacenter.

Every number real and read today: the two options differ by two orders of magnitude and the user cannot choose without seeing that. Wait for the pick, and do not relitigate it afterwards.

Three exits at this step:

• **The model is not in the OVH catalog** (DeepSeek, GLM class): say so in one line; the recommendation collapses to RunPod alone.
• **Their words already picked**: "pay per use" is Route A; a single-tenancy requirement is Route B. Skip the menu.
• **The double requirement**, EU ownership and single tenancy at once: neither option qualifies. Name Hetzner, Verda or Scaleway and stop rather than fudge.

4. The token

Guide them to the credential for the provider they picked, and only that one. **Give the exact URL as a clickable link, never just the click path.** Both verified 6 August 2026:

• **OVH**: https://www.ovh.com/manager/ (redirects to their regional manager) → Public Cloud → AI & Machine Learning → AI Endpoints → API keys. The keyless trial at 2 requests/min can prove the route before they