Your AI agent, around the clock
Put your AI agent on a VPS and it keeps working even when you close your laptop or shut it down. Claude Code, coding agents and n8n automations run on the server 24/7, not on your computer.
Your laptop can shut down. The VPS keeps running
A VPS is a remote Linux server that stays powered on continuously, so an AI agent launched there keeps running without depending on your personal computer staying awake.
Your agent inside your laptop
- You close the lid and the process stops
- A reboot or a power cut and it does not come back on its own
- You switch networks, the IP changes and webhooks stop arriving
- It depends on your battery, your wifi and your schedule
Your agent inside a CubePath VPS
- Powered on continuously, with redundant power and network
- Your own static IP for your webhooks, your bots and your SSH
- Full root to install Docker, Python or Node
- The process keeps going once you log out, under systemd or in a container
- Our own AS26141 network with DDoS protection included
What server does an AI agent need?
It mostly depends on where the model runs. What you will see below are recommended starting configurations, not requirements: real usage is set by whatever your agent runs.
The model lives in an external API
Agents that call the model through an external API, such as Anthropic's or OpenAI's, normally do not need a GPU on the VPS: model inference happens on the provider's infrastructure. Your server handles everything else.
- The agent runtime and your own code
- Docker containers and their dependencies
- Database, queues and the agent's memory
- Builds, tests and a headless browser
Ordinary vCPU, RAM and disk are usually enough. That is where the general purpose range starts, from $4.06 a month.
You want the model inside the server
If you run the model itself locally, with Ollama or similar, the requirements are different and they are set by the size of the model, not by the agent.
- Small quantised models: a lot more RAM and CPU
- Large models: usually ask for GPU acceleration
- Disk for each model's weights
- Latency and cost different from an API
For that, take a look at self-hosted AI or the AI Gateway, which lets you use several models with a single key and a single bill.
Recommended configurations to start with
Three common starting points. You can start with the smallest and resize when the agent grows, because billing is hourly.
gp.nano
- 1 vCPU
- 2 GB RAM
- 40 GB NVMe
- 3 TB transfer
An agent that calls the model over an API, with cron and alerts to Telegram or Slack.
Deploy this VPSgp.micro
- 2 vCPU
- 4 GB RAM
- 80 GB NVMe
- 5 TB transfer
Docker, n8n and a coding agent living together comfortably.
Deploy this VPSgp.starter
- 4 vCPU
- 8 GB RAM
- 100 GB NVMe
- 10 TB transfer
Several agents or workers, a headless browser and your own database.
Deploy this VPSWhat real usage depends on
Two agents calling the same API can use very different resources depending on what they do between calls. Start with a small configuration, watch the usage and resize.
| What you are going to run | Typical configuration | Resources | From | Deploy |
|---|---|---|---|---|
| A light agent, cron and Telegram alerts | gp.nano | 1 vCPU / 2 GB | $4.06/mo | Deploy |
| n8n with Docker and a couple of agents together | gp.micro | 2 vCPU / 4 GB | $8.11/mo | Deploy |
| Several agents, headless browser and a database | gp.starter | 4 vCPU / 8 GB | $15.21/mo | Deploy |
| Builds, tests and several workers in parallel | gp.small | 8 vCPU / 16 GB | $29.40/mo | Deploy |
| Small models running locally with Ollama | gp.medium | 12 vCPU / 32 GB | $55.76/mo | Deploy |
Monthly compute prices (hourly rate × 730). The dedicated IPv4 costs an extra $1.50/mo. General purpose plans go up to 48 vCPU and 192 GB of RAM. See all plans and prices or everything about the VPS.
How an agent that never shuts down works
The agent runs on the server, not in your session. That is why it keeps working when you close the laptop.
You connect over SSH whenever you want, give instructions and disconnect. The process belongs to the server, so it keeps going with the lid closed, with your phone in your pocket or while you are on holiday.
Underneath it is a plain, ordinary Linux. You install Python, Node or Docker and leave the agent under systemd or inside a container with automatic restart.
The IP does not change, so webhooks from GitHub, Telegram or your CRM keep arriving at the same place months later. When you come back, what you find is work already done.
Your laptop
You only use it to connect when you want to. You can close it at any moment.
Your CubePath VPS
On 24 hours a day in Barcelona, Houston or Miami, with its own static IP.
Docker or the agent runtime
Containers, Python or Node environments and systemd processes that come back up on their own.
Your AI agent
The loop that decides, executes and waits for the next event without interruptions.
APIs, GitHub, databases
It calls the model, makes commits, queries data and pings you on Telegram or Slack.
Everything below the first layer keeps running even if you turn the laptop off.
What your agent can do while you are away
Continuous or long-running tasks that only make sense if they never stop halfway.
Let it keep coding with the laptop closed
Claude Code, other coding agents or your own custom one all need the same thing: a machine that will not fall asleep halfway through the task.
Leave the agent working and close the laptop
You launch the session inside tmux or as a service and the process belongs to the server, not to your terminal. You come back hours later and read what it has done.
Background work on your repository
Migrations, tests, dependency updates or long refactors that end up in a branch waiting for you.
Builds and tests without tying up your machine
Compilations and long suites run on the server, with its vCPU and its NVMe disk, while you keep working locally.
Jobs scheduled for the early hours
A cron at 03:00 launches the agent, reviews whatever needs reviewing and leaves the report ready for when you wake up.
Several agents or workers at once
One container per agent or per project, isolated from each other, sharing the same server and the same bill.
Agents that are always reachable
A GitHub webhook or a Slack message wakes the agent up because the IP does not change and the port stays open.
Workflows, alerts and monitoring without pauses
Automations with n8n
Workflows running without interruptions, with webhooks always reachable on a static IP that does not change.
Email triage
It checks the inbox every few minutes, summarises what matters and pings you on Telegram or Slack only when it is worth it.
Monitoring and watching
It watches prices, sites or availability 24 hours a day and reacts the exact minute something changes.
Memory that outlives the session
Postgres, Redis or a vector database on the same server, with NVMe disk and snapshots before every big change.
Your agent can manage the infrastructure too
CubePath publishes a REST API and an open source CLI that ships MCP servers. An MCP-compatible agent can create a server for a test, resize it and shut it down when it finishes, without you touching the panel.
Claude Code and Anthropic, OpenAI, n8n, Docker, Coolify, Dokploy and Ollama are trademarks of their respective owners. CubePath provides servers to run that software on and holds no official partnership or integration with those companies.
Why CubePath for AI agents
This is not the spec sheet. It is what each spec means when the process has to last for months.
Redundant power and network
Your agent runs on AS26141, CubePath's own network, with DDoS protection included. An agent that goes down halfway through a task loses the context of what it was doing.
Barcelona, Houston and Miami
You choose where the agent lives: Spain if you need the data here, the United States if your agent mostly talks to APIs hosted there.
Full root on Linux
Ubuntu, Debian, AlmaLinux, Rocky Linux or Fedora with root access. You install Docker, Python, Node or whatever your agent asks for, with no sandbox and no blocked libraries.
Snapshots and automatic backups
An agent with write permissions can break things. You take a snapshot before giving it more freedom and schedule automatic backups with up to 7 days of retention.
REST API and CLI with MCP servers
Your own agent can create, resize or power off servers by itself: the CubePath CLI is open source and ships MCP servers.
Hourly billing, no commitment
You try an agent for an afternoon and pay for the hours it was on. If it grows, you scale up to 48 vCPU and 192 GB and come back down once the peak passes.
The same server holds the rest of your stack
An agent rarely lives alone. It ends up surrounded by containers, databases, APIs and a panel to deploy it all, and that fits whole on the same machine.
n8n
AI workflows, webhooks and cron.
Ollama
Open models running on your own server.
Coolify & Dokploy
Panels to deploy your projects on Docker.
Portainer
Visual management of the server's containers.
Supabase
Database and auth for whatever your agent builds.
Uptime Kuma
Checks that your agent is still alive and tells you if it is not.
All of them deploy in 1 click from the CubePath marketplace, or you install them by hand: the server is yours and you have root.
Frequently asked questions
What people ask before moving their agent to a server.
Can I run an AI agent 24/7 on a VPS?
Yes. A VPS is a remote Linux server that stays powered on continuously, so an agent launched there keeps running even if you turn your computer off, close your laptop or lose your connection. That is the difference between a script that runs while you watch it and an agent that works on its own.
What is the difference between running an agent on my laptop and on a VPS?
On a laptop the agent depends on your session: close the lid, reboot or switch networks and the process stops and the IP changes. On a VPS the agent lives on a remote machine that is always on, with a static IP, so webhooks keep arriving and long tasks finish even when you are not around.
Do I need a GPU to run AI agents?
Usually not. Agents that call the model through an external API, such as Anthropic's or OpenAI's, do not need a GPU on the VPS: inference happens on the model provider's infrastructure and the server only runs the agent runtime, its code, its containers and its database. A GPU or a lot more RAM comes into play if you want to run the model itself inside the server.
Can I leave a coding agent like Claude Code working with the laptop closed?
Yes. You connect over SSH, launch the agent inside tmux, screen or a systemd service and close the session: the process belongs to the server and not to your terminal. When you come back you reconnect and see everything it did while you were away. CubePath provides the servers to run it on and is not affiliated with Anthropic or any other model provider.
Which plan should I start with?
As a starting point, gp.nano (1 vCPU and 2 GB) is usually enough for a light agent with cron and alerts; gp.micro (2 vCPU and 4 GB) is comfortable with Docker, n8n and development; and gp.starter (4 vCPU and 8 GB) leaves room for several agents, a headless browser and your own database. These are recommendations, not requirements: real usage depends on the tools you run, the number of agents, the builds and tests and the containers you bring up. You can start small and resize later.
How many agents can I run on the same server?
There is no per-plan limit: you can fit as many as the CPU, RAM and disk you have contracted allow. The usual approach is to isolate each agent in its own Docker container and move up a plan when the server gets tight, since resizing is done from the panel or from the API.
Can I use Docker and deploy n8n?
Yes. You get full root access on Ubuntu, Debian, AlmaLinux, Rocky Linux or Fedora, so you install Docker and whatever else you need. n8n, Ollama, Coolify, Dokploy, Portainer and Supabase are also in the CubePath marketplace as 1-click deployments.
What happens if the server reboots?
The server boots back up and, if you left the agent as a systemd service or inside a container with a restart policy, the process comes up with it. You can also create a snapshot before any big change and schedule automatic backups with up to 7 days of retention.
Can I change plan when my agent grows?
Yes. General purpose plans go from 1 vCPU and 2 GB up to 48 vCPU and 192 GB, and the server is resized from the panel or from the API. Since billing is hourly, moving up a plan for a busy stretch and back down afterwards costs little.
Where are the servers?
CubePath deploys VPS in Barcelona (Spain), Houston (Texas) and Miami (Florida) on its own AS26141 network. You choose the location at deploy time, based on where you prefer to keep the data and the latency.
How long does it take to be ready?
Under 30 seconds. You choose location, plan and operating system, and you get the IP and root access to launch your agent right away.
Can the agent itself manage the infrastructure?
Yes. CubePath publishes a REST API and an open source CLI that ships MCP servers, so an MCP-compatible agent can create, resize or power off servers on its own with the credentials you hand it.
Still have one? Write to us or drop by the general questions.
Give your AI agents a server that never shuts down
Ready in under 30 seconds, no commitment and hourly billing.