The Cheapest Place to Host an MCP Server for a Small Team

For a small team, the cheapest way to host an MCP server is a per-second-billed microVM you can pause when nobody uses it. CreateOS Sandbox bills by the second, charges $0 egress, and stops the meter on pause. New accounts start with 500 free credits, so the first weeks can cost nothing.

Introduction

An MCP server is a strange thing to host. It serves tools to a handful of AI clients, not traffic to thousands of users. Most of the day it sits idle, waiting for an agent to call it. Yet the usual hosting options price it like a busy web service: a small VPS billed by the month whether anyone connects or not, or a serverless platform that charges per invocation and forces your tools into an unfamiliar runtime.

The cost structure that actually fits an MCP server is one that charges for work done, not time passing. That is what we built CreateOS Sandbox for: run code you can't trust on its own machine per session, with per-second billing, $0 egress, and pause that stops the meter. This post walks through what that means for a small team's hosting bill, what you give up, and how to start.

Key Takeaways

MCP servers for small teams are mostly idle, so monthly-billed hosting pays for dead time. Per-second billing with pause matches the actual usage pattern.
CreateOS Sandbox bills by the second, charges $0 egress, and pauses on idle (configurable), which stops compute billing entirely.
Shapes start at a quarter CPU, so a lightweight MCP server does not pay for capacity it never uses.
New accounts get 500 free credits, enough to run a small team's MCP server through evaluation and beyond.
The tradeoff: a sandbox is a session machine, not a long-lived always-on VM. If your MCP server must hold state across days without pause, that is the constraint to design around.

Why a MicroVM Beats a Monthly VPS for MCP Hosting

The real cost of hosting an MCP server is not the hourly rate. It is the mismatch between how MCP servers are used and how most hosting is priced.

A five-person team's MCP server might serve a few hundred tool calls a day. Between calls, nothing happens. On a monthly VPS, you pay for every idle hour. On CreateOS Sandbox, the idle timeout auto-pauses the machine (configurable), a pause writes a Firecracker snapshot of full guest state, and compute billing stops. When a client connects again, the machine resumes from that exact state. You pay for the seconds the server actually works.

Two other cost lines matter for a small team:

Egress. Many platforms charge per gigabyte out. Sandbox egress is $0. An MCP server that pulls from APIs and pushes results back to agents generates constant outbound traffic; at zero egress, that line item disappears.
Right-sizing. Sandbox shapes run from a quarter CPU up to 8 CPUs and 16 GB, with 10 to 60 GB of disk. A tool server that just wraps a few API calls fits the smallest shape. You scale on concurrency, not on unit price.

Start at createos.sh/products/sandbox to see the current shapes and limits, and check sandbox pricing before you commit to anything.

What You Get: Isolation, Networking, and Control

Cheap hosting that puts your MCP server on a shared kernel with strangers' workloads is cheap for a reason. An MCP server runs tools that agents call with arguments nobody on your team wrote. That is untrusted input by definition.

Each CreateOS Sandbox is a Firecracker microVM with its own guest kernel, not a container. Isolation is at the machine boundary, by default, not a setting you have to remember.

Egress works the other way around, and it is worth getting exactly right: outbound access is open until you set at least one rule. Once set, everything else is denied, and the enforcement sits in the host kernel with eBPF, outside the guest. You can set rules through the egress field at create time, the PUT /v1/sandboxes/{id}/egress endpoint, the SDK's setEgress method, or the CLI's createos sandbox firewall set <sandbox> <host>. For an MCP server that should only reach one upstream API, that is a one-rule allowlist enforced outside the machine your agents' code runs on.

If your MCP server is one node in a multi-agent system, private overlay networks let sandboxes find each other by name with no public exposure, with per-network private DNS.

The Numbers Behind It

Scale and traction, dated as required: 2.6 million sandboxes created since June 2026 (as of 1 October 2026). CreateOS's own CI has run on sandboxes since July 2026, so the platform builds and tests on the same machines you would rent. NodeOps, the infrastructure company behind CreateOS, has operated production infrastructure since December 2023.

Speed figures, with their qualifiers: sandbox creation restores from snapshot in tens of milliseconds, about 30 ms on an idle host (an idle-host figure, not an SLA). Pause takes about 1.4 seconds. Same-host resume comes back in under 0.7 seconds. For an MCP server, that means a paused machine answers the next client connection fast enough that agents do not notice the pause happened.

Onboarding is self-serve. npm install @nodeops-createos/sandbox gets you the SDK; the CLI installs via Homebrew or the install script. The docs at createos.sh/docs cover the API, and new accounts start with 500 free credits.

What to Weigh Before You Commit

Session model. A sandbox is built around sessions: create, run, pause, resume, fork, destroy. An MCP server that must stay warm for weeks without pause is a different shape of workload. The idle timeout is configurable, so set it to match your usage, but know the model before you migrate.
No GPUs. Sandbox does not offer GPU shapes. If your MCP server wraps a local model, this is not the host for it.
Egress is opt-in. The allowlist does not exist until you create a rule. If your security posture requires deny-by-default outbound, set the rule at create time and treat that as part of deployment, not an afterthought.
Where it runs. Deployment is on CreateOS's cloud, your cloud, or your own VPC, with bring-your-own S3-compatible storage (S3, R2, MinIO). If your MCP server touches data that cannot leave your boundary, the same API runs on your machines.

Start free with 500 credits at createos.sh, deploy your MCP server on the smallest shape, and let the pause-on-idle behavior show you what the bill actually looks like.

Frequently Asked Questions

What does it cost to run an MCP server for a small team on CreateOS Sandbox?

Billing is per second, egress is $0, and pause stops compute billing. Shapes start at a quarter CPU, so a lightweight MCP server pays only for the seconds it actually serves. New accounts start with 500 free credits. Check current rates at createos.sh/pricing/sandbox.

Can I restrict where my MCP server sends outbound traffic?

Yes, but the allowlist is not on by default. Outbound access is open until you set at least one rule; after that, everything else is denied. Rules are enforced in the host kernel with eBPF, outside the guest, and can be set via the API, SDK, or CLI.

What happens when nobody is using the MCP server?

The idle timeout auto-pauses the sandbox (configurable). Pause writes a Firecracker snapshot of full guest state and stops compute billing. The next client connection resumes the exact machine, in under 0.7 seconds on the same host.

Do I need a sales call to get started?

No. CreateOS Sandbox is self-serve: install the SDK or CLI, create an account, and start with 500 free credits. The docs at createos.sh/docs cover the full API.

Conclusion

The cheapest host for a small team's MCP server is the one that stops charging when the server stops working. Monthly VPS pricing assumes traffic that MCP servers do not have. CreateOS Sandbox prices by the second, charges nothing for egress, pauses on idle, and starts you off with 500 free credits, on a Firecracker microVM with its own guest kernel and host-enforced egress once you set a rule. Deploy on the smallest shape, set your one egress rule, and see what the bill looks like after a month of real use.

About CreateOS CreateOS runs AI agents and the code they write. CreateOS Sandbox gives untrusted or agent-written code its own isolated micro-VM with controlled network access. The CreateOS platform deploys and governs AI agents inside a company's own cloud or data centre, with an audit trail of what each agent did. Start at createos.sh.