Glossary

Active-CPU billing: The processor is charged only while it is working, not while it waits. Cheap for agents that spend most of their time idle. Read more
API and SDK: Ways for a program, rather than a person, to control the provider: create a machine, run a command, delete it. An SDK is a ready-made library for one programming language. Read more
Cache read: Tokens the model has already seen in this conversation and can reread cheaply. Agents reread a lot, so this is most of their tokens and much cheaper than new input.
Cold start: The wait while a machine is created or woken before it can do anything. From under a second to a minute, depending on the provider.
Custom domain: Serving what the machine runs at your own web address, such as agent.example.com, instead of the provider's. Read more
Disk: Where files are kept. On some providers the disk survives a stop, on others it is wiped, which is why the guide asks whether files must be kept. Read more
Egress: Data the machine sends out to the internet, such as downloads to a user or uploads to another service. Some providers charge per GB; most agents send very little. Read more
Fork: Starting one or more new machines from a snapshot, each continuing from the same point. Useful for trying several things in parallel. Read more
GiB / GB: Units of memory or disk. A GiB is about 7% larger than a GB; for choosing a provider the difference rarely matters.
Idle timeout: How long a machine can sit unused before the provider pauses, stops or deletes it. Short timeouts save money but can break an agent that must answer at any time. Read more
Included usage: A monthly plan fee that comes with the same amount, or a set amount, of usage. You pay the fee even if you use less, and pay extra only above it. Read more
Isolation: How firmly one machine is kept apart from others on the same hardware. A microVM is a small full computer and is the stronger wall; a container shares more with its neighbours. Read more
MCP: Model Context Protocol: a common way to give an AI agent tools, such as reading a calendar or searching a database, served by a small program called an MCP server.
Memory (RAM): The machine's working space while it runs. Lost when the machine stops unless the provider saves it. A browser or a large program needs more of it. Read more
Per-second billing: Usage is counted to the second, so a two-minute task costs two minutes. Some providers round up to a minimum first. Read more
Region: The part of the world where the machine runs. It matters for speed if your users are far away, and for law if data must stay in a country or the EU. Read more
Sandbox: A computer you rent in the cloud, sealed off from everything else, that a program can create, use and throw away. Agents use one so the code they run cannot reach your own machine or other customers. Read more
Sleep and wake: The provider pauses an idle machine and starts it again when a request arrives. You pay little while it sleeps, but the first reply after a pause is slower. Read more
Snapshot: A saved copy of a machine at one moment, so it can be restored or copied later. Some providers save memory and running programs too, others only files. Read more
SSH: A standard way to open a terminal on a remote machine from your own computer. Read more
Terminal: A text window where you type commands instead of clicking. Many providers can be set up only this way; the guide says when one cannot.
Token: The unit an AI model is billed in, roughly three-quarters of a word. An agent re-reads its instructions and history on every message, so one message can use tens of thousands.
vCPU: One share of a processor. More vCPUs let a machine do more things at once; an agent waiting for messages uses a small fraction of one. Read more
VPS: A virtual private server: a machine you rent by the month that runs all the time. Cheaper for one always-on agent, but you set it up and look after it yourself. Read more
Webhook: A web address that another service calls when something happens, such as a new chat message. It lets a sleeping machine be woken only when needed.
Zero data retention: The model provider promises not to store your prompts or replies after answering. This guide prices only model offerings that make this promise.