Choosing a VPS for self-hosted AI: run Hermes Agent, Claude Code, n8n, Open WebUI and chatbots, on a yearly plan in a region Claude supports (2026-10)

How much server you need

Anthropic's official minimum for Claude Code is 4GB RAM, an x64 or ARM64 processor, and Debian 10 or Ubuntu 20.04 or later; on large repositories a single session can use more than 2.5GB. An agent does not just call an API, it works on the machine: Claude Code clones repositories, installs dependencies, runs builds and tests, and opens a headless browser to look at pages, all of which use memory; with 4G, a browser open plus one build is enough to run out. We recommend 2 vCPU 8G: the official 4G is the floor, and 8G lets the agent, the browser and builds run at the same time without crowding each other; for a light agent that only calls APIs, such as Hermes Agent, 4G will do; for n8n and multi-service Docker setups, also plan on 8G. Disk 40G or more: Docker images, node_modules and browsers each take a few GB. Traffic is small, only API requests and text go back and forth, so there is no need to pay extra for traffic.

Browser agents need more memory. Agents like OpenClaw that open Chrome and click through web pages themselves, like Claude Code running Playwright tests, use a few hundred MB per tab; a session with several pages open takes one or two GB, and with the agent itself and the system on top, that is the main reason for 8G rather than 4G. To run two such agents at the same time, buy 16G.

Put the server in a region Claude supports

Requests to the model API go out from whichever machine the agent is installed on. Anthropic only provides service in the countries and regions it supports, and neither mainland China nor Hong Kong is among them, so the server must be in a supported region such as the US, Japan or Singapore. The “Claude” column in this site's tables follows the official list of supported regions; whether you can use it is determined by Anthropic's official list and terms of use.

If you are in China and often open the agent's web panel, the server's route affects how responsive the panel feels, but not the agent's own work. For choosing a route, see How to choose a China route for Telecom, Unicom and Mobile.

8G plans in stock: available for Claude, sorted by effective monthly price

An agent runs around the clock, so just pay yearly: the effective monthly price is often half the monthly rate, the server stays put, and you do not have to deal with renewals every month. In the table below, “Monthly price” is the yearly price divided by 12 where a yearly price exists, converted to US dollars; the “List price and billing” column is the vendor's list price and billing period, and plans whose first-year price differs from the renewal price are marked “Renewal”. Yearly plans are often limited stock; if one is out of stock, wait for a restock. This site checks stock daily. The cheapest right now is $8.99/mo. 8G is this site's Main server tier; filter “Use” by “Main server” on the home page to see them all. Click a column header to sort; for the full list and filters, see the home page.

How big a server Open WebUI, Dify, n8n and AnythingLLM need

These are self-hosted AI suites: Open WebUI and AnythingLLM are your own chat interface, Dify builds workflows and knowledge bases, and n8n is workflow automation with LLM nodes. They do not run models themselves, they only pass requests on to the Anthropic or OpenAI API, so they need memory rather than compute: Open WebUI or n8n alone runs fine on 2 vCPU 4G; Dify brings up a dozen or so containers (PostgreSQL, Redis, a vector store, a sandbox), so plan on 8G; more if the knowledge base computes embeddings locally. The data centre must also be in a region Claude supports, because the API requests go out from this machine. The 8G table above is meant for these too; no need to buy anything else.

What server to use for chatbots and MCP services

A Telegram, Discord or Feishu bot connected to an LLM, or an always-on MCP service or webhook receiver, only receives messages, calls an API and replies; idle memory use is a few dozen MB, so a smallest plan of 1 to 2G is enough, the cheapest tier. The requirements are the same as above: the server in a region Claude supports, running 24 hours a day, paid yearly. The table below shows the cheapest plans that are in stock now, available for Claude, with 1 to 2G, sorted by monthly price; click a column header to sort. Most of these plans come with China-optimised routes; bots do not need the route, but it does no harm.

Can you run open-source models on a VPS

Not recommended. Running a 7B model in Ollama takes 8 to 16G of RAM, and a VPS without a GPU produces only a few tokens per second, a far worse experience than calling an API, and no cheaper; to run models you need to rent a GPU cloud, which is not a VPS and is not listed on this site. The right uses for a VPS are the ones above: running agents, suites and bots, and leaving the model to the API.

How to deploy: install Debian and let Claude Code or Codex on your own computer do the rest

1. Choose Debian as the OS. It has the smallest footprint, so the same RAM leaves more room for the agent and the browser, which amounts to saving money; it is usually among the vendor's images.
2. You do not need to type commands yourself. Open Claude Code or Codex on your own computer, give it the VPS address and key, and let it SSH in to install Docker, pull the agent and set it to start on boot; when it is done, have it write down for you how to access it.
3. Put the API key in an environment variable or a .env file; never commit it to the code repository.
4. Do not expose the agent's web panel directly to the internet: have it listen on localhost only and reach it through an SSH tunnel, ssh -L 5678:127.0.0.1:5678 user@server-ip; or put a password-protected reverse proxy in front of it.
5. Open only the SSH port in the firewall, log in with a key and disable password login. Use systemd or Docker's --restart unless-stopped so services restart automatically after a crash.