AI agent hosting selector

1. The hosting category

Only GPU need, uptime and management decide the category. The rules are checked top to bottom and the first match wins.

1. GPU needed and the agent runs 24/7Always-on GPU instanceCompare GPU cloud providers
2. GPU needed and the agent runs on demandServerless GPU inferenceCompare inference hosting providers
3. No GPU and you want a managed platformManaged agent platformCompare managed AI hosting
4. No GPU and the agent runs on demandServerless agent platformCompare AI agent hosting
5. No GPU and the agent runs 24/7 on a server you controlAI VPS (“with persistent storage” when storage is required)Compare AI VPS providers

2. Ranking

1. Runtime support and workload fit are always required. Management is required when you pick managed or self-managed, and the budget whenever you set a limit. The other questions are Required, Preferred or Not needed.
2. Each check scores 2 points when met, 1 when partly met and 0 when not verified or not met. Required checks count 2 times.
3. A provider that does not meet a required check is ruled out and listed with the reason. A check we could not verify never rules a provider out: it scores nothing and shows up as a trade-off.
4. The budget of an agent that runs 24/7 is the exception: only a price we can check passes it. A provider whose price is not published, or not republished under its terms, is neither ranked nor ruled out. It is named next to the shortlist with a link to its pricing, so it is never the answer to your budget.
5. Providers that run the workload come first, then those that run it only partly (for a GPU agent on demand: GPU instances you start and stop yourself, or a per-token model API), so a partial fit never outranks a full one. Within each group, providers are sorted by points, then by runtime support (“Official template”, then “Documented setup”, then “Community guide”; “Manual setup”, a model API or no record rank equal), then by entry price (lowest first, unpriced last), then by name. The first 3 form the shortlist.
6. Only hosting counts as runtime support. A provider whose models the runtime calls (a model API, for example for Hermes Agent) or that only integrates with the runtime is shown as not verified, and so is “Manual setup” on a generic server.
7. A serverless product only counts as running your agent when the provider publishes serverless GPU prices billed while your containers run. A per-token model API cannot run your agent or your own model, so it counts as partly, with the provider's own description shown. For an agent without a GPU, the check says that these published prices are for GPU containers.
8. The EU and US checks read where the offered plan or GPU runs, never the provider's other products: a provider's EU servers are no evidence for its managed agent plan. With a GPU needed, they read the regions of the GPU offers themselves, the way the GPU cost calculator does; for a 24/7 agent without a GPU, the region scope of the always-on plans. The plan or GPU behind the entry price must be listed in the area to meet. One whose location there is not published or not recorded is not verified, and one that runs only elsewhere does not meet it. With a location required, only plans and GPUs that can run there set the entry price; with it preferred, the cheapest one listed there sets it when it fits your budget, and the check counts as partly when only one elsewhere does. The US means the country, not North America.
9. “Managed platform” means the provider runs the agent's servers: a managed agent plan, or a serverless platform for GPU or on-demand agents. Managed Kubernetes, databases or model APIs next to a server you run yourself count as partly.
10. Entry price: for an agent without a GPU that runs 24/7, the cheapest always-on server plan with at least 1 GB of RAM (OpenClaw's own DigitalOcean guide runs on 1 GB), at its standard or renewal monthly price, or its hourly price × 730 hours. Introductory prices are never used. If you ask for a managed platform, a provider with a managed agent plan is priced by that plan. With EU or US required, only plans that can run there are used, so a price from one region never stands in for another. With a GPU, the cheapest no-commitment GPU offer × its GPU count × 730 hours, or the provider's monthly price where its hourly billing stops at one (Hetzner); spot prices can be interrupted and reserved prices need a term, so neither counts. One-off setup fees are named next to the price, not added to it. On-demand workloads have no entry price, because the cost depends on how long the agent runs.
11. Prices a provider's terms do not let us republish are never used: that provider has no entry price and is never checked against a budget. With a budget limit for a 24/7 agent it is left out of the ranking and named next to the shortlist with a link to its live pricing; on demand or without a limit it is ranked on its other checks, and its card links to its live pricing.
12. The next steps link to the category comparison, repeat your required filters, then point to a runtime compatibility page (or the VRAM calculator when you need a GPU, set to Hermes Agent's 64K minimum context for Hermes) and our guide to running an agent 24/7.