a2a cloud
pay when it runs

Scale-to-zero AI agent hosting.

An AI agent is idle most of the day, but a rented box bills every hour — and that hosting cost stacks on top of per-token model spend. Scale-to-zero fixes the #1 complaint about running agents: idle plus per-token cost. On a2a cloud each agent runs in a libkrun microVM that scales to zero when there's no traffic and wakes on the next request. You pay for invocations, not for uptime. Idle costs nothing.

idle = free · pay per invocation · free tier

0$
idle cost
0
free invocations/mo
0
free agent
the problem

You're paying for an agent that mostly waits.

Most agents spend the overwhelming majority of their life idle — waiting for the next request. Host one on a container or VM and you pay for every one of those idle hours anyway, at capacity sized for a peak you rarely hit. Layer per-token model spend on top and idle becomes the most expensive part of running an agent: you're billed to keep a box warm so it can do nothing.

A rented container or VM bills 24/7 regardless of how many requests it serves.
Capacity is provisioned for peak traffic, so you overpay during the long quiet stretches.
Hosting cost stacks on top of per-token model spend — idle is billed twice over.
Real isolation usually means a dedicated VM you keep warm and pay for continuously.
the a2a way

Right-size the agent to the traffic it actually gets.

a2a runs every agent — and the database, MCP server, gateway, and frontend that ship with it — in libkrun microVMs that scale to zero. The whole agent app is billed against real traffic, not provisioned for a peak that's idle most of the time.

Scale to zero when idlePay per invocationFree tier to startFast cold startsIsolation without the idle billThe whole app scales with it

Scale to zero when idle

Each agent runs in a libkrun microVM that shuts down when there's no traffic and wakes on the next request. An idle agent holds no compute — so it costs nothing to keep deployed.

Pay per invocation

Billing follows requests, not wall-clock. You pay for the runs your agent actually serves instead of renting a box by the hour whether it works or sleeps.

Free tier to start

The free tier covers one hosted agent and 10,000 invocations a month. Ship a real agent, put it in front of users, and pay nothing until you outgrow it.

Fast cold starts

libkrun microVMs boot far faster than a full VM, so scale-to-zero doesn't mean a painful wait. The agent wakes, serves the request, and idles back down.

Isolation without the idle bill

Each microVM is isolated from the host and other tenants. You get real isolation for untrusted agent code without paying to keep an isolated box warm around the clock.

The whole app scales with it

The managed Postgres, MCP server, gateway, and frontend that ship with the agent ride the same model — so the whole agent app is right-sized to actual traffic, not provisioned for peak.

side-by-side

Always-on box vs. scale-to-zero.

dimension
rented vm / container
a2a cloud
idle cost
A rented container or VM bills 24/7 even when the agent serves zero requests.
Scale-to-zero microVM holds no compute when idle — you pay nothing for downtime.
billing model
You pay by the hour for provisioned capacity, sized for peak you rarely hit.
You pay per invocation — cost tracks the runs your agent actually serves.
on top of tokens
Hosting bills stack on top of per-token model spend, doubling the cost of idle.
Idle hosting is free, so per-token spend is the only cost while the agent waits.
isolation
Real isolation usually means a dedicated VM you keep warm and pay for continuously.
libkrun microVM isolates untrusted code without an always-on bill.
getting started
Even a demo agent needs a paid instance running before anyone can try it.
Free tier: one agent and 10,000 invocations a month, deployed at no cost.
questions

Frequently asked.

What is scale-to-zero agent hosting?

Scale-to-zero means your agent runs only when it's handling a request. On a2a cloud each agent lives in a libkrun microVM that shuts down when idle and wakes on the next call. Because an idle agent holds no compute, keeping it deployed costs nothing — you pay for invocations, not for uptime.

How does pay-per-invocation lower agent hosting costs?

Agents are idle most of the time, but a rented container or VM bills around the clock — on top of per-token model spend. Pay-per-invocation ties cost to the runs you actually serve, so idle time is free and hosting no longer doubles the cost of an agent that mostly waits.

Doesn't scale-to-zero mean slow cold starts?

libkrun microVMs boot far faster than a full virtual machine, so waking from zero is quick. The agent starts, serves the request, and idles back down — you get the cost profile of serverless without the long cold-start penalty people associate with it.

Is there a free tier for hosting an AI agent?

Yes. The free tier includes one hosted agent and 10,000 invocations per month, so you can deploy a real agent — with its database, MCP server, gateway, and frontend — and put it in front of users before paying anything.

Do I still get isolation with scale-to-zero?

Yes. Each agent runs in its own libkrun microVM, isolated from the host and other tenants, whether it's serving traffic or scaled to zero. You get real isolation for untrusted agent code without paying to keep an isolated box warm 24/7.

keep reading

Related guides.

All guides live in the guides index.

idle costs nothing

Pay only when your agent runs.

a2a cloud deploys any agent — LangGraph, CrewAI, AutoGen, the OpenAI Agents SDK, or custom — into a libkrun microVM that scales to zero when idle, with a managed Postgres, an MCP server, an API gateway, a hosted frontend, and an Ed25519-signed receipt for every run. Start on the free tier: one agent, 10,000 invocations a month, the whole agent app, no idle bill.