How many agents fit on Pro-32 vs Pro-64?

No clean number — it depends how much each agent keeps in memory. Rough feel: a crew of 5–10 with shared memory plus a modest vector store is comfortable on 32 GB; a couple dozen agents, bigger histories, or a large memory index moves you to 64 GB. Start lower, resize when you hit swap.

My agents just call an LLM API — do I need this?

Probably not. Light, API-bound agents barely touch local memory; a NAT or dedicated-IP plan from $3–20 handles them fine. Pro earns its place once agents hold shared state — history, a working set they pass around, a vector store for memory.

Do I need a GPU for an agent fleet?

No. Agents orchestrate and call an LLM; the model runs on your provider's GPUs. Locally you need CPU and RAM for coordination, memory and any vector search. That's exactly what a Pro plan is.

Why no-KYC for an agent system?

If an agent provisions its own infrastructure, or the workload touches data you'd rather not tie to your identity, paying in crypto with no ID keeps the whole loop private — and an agent can order the server itself over MCP, no human step.

VPS for multi-agent systems and agent fleets

One agent is cheap to host. A crew of ten holding shared state is a memory problem. Sizing a CrewAI/LangGraph fleet, why state eats RAM, and where no-KYC high-memory fits. From $40/mo.

The most common mistake with agent hosting is sizing for one agent. Someone runs a single agent, watches it use 400 MB, and concludes agents are cheap to host. Then they scale to a real crew and the box starts swapping at 3 a.m.

One agent is cheap. That's not the interesting case.

Where the memory actually goes

An agent that only fires API calls at a model is light — it's mostly waiting on the network. Run a dozen of those on a small plan and you'd never notice.

The RAM disappears when agents hold state. History that grows every turn. A working set several agents read and write. A vector store for long-term memory in the same process. The moment the architecture stops being "call API, forget" and becomes "remember, coordinate, hand off," memory is the constraint — not CPU. CrewAI, LangGraph, AutoGPT-style loops all trend this way as they get serious. The framework doesn't eat the RAM; the state does. We dug into the numbers here.

Rough sizing

Light, API-bound agents — you don't need Pro; a NAT or dedicated-IP plan ($3–20) is plenty. Skip the rest of this page.
A real crew (5–10 agents) with shared memory plus a useful vector store — Pro-32 (32 GB) is the sweet spot. Most fleets land here.
Bigger fleets, longer histories, a memory index in the millions — Pro-64 (64 GB). This is where one box replaces the three smaller ones you'd otherwise juggle.
Fleet plus co-located services, or agents plus local inference — up to Pro-80 (80 GB).

Start below where you think you need to be, watch htop for a day, resize up when you see swap. Guessing high just wastes money.

Why this shape of host

A fleet that provisions or manages its own servers wants an API it can drive without a human — and increasingly, to pay without one too. High memory, a dedicated IP, crypto payment, no KYC, and the whole thing orderable by an agent over MCP: that combination is rare, and it's what the Pro line is built for. It's not the cheapest per gigabyte, and if your agents are light you genuinely don't need it. But for a serious system that holds state and values privacy, it's the right fit.

When you're there, Pro-32 covers a real crew; size up to 64 or 80 GB as the fleet grows.

VPS for multi-agent systems and agent fleets

Where the memory actually goes

Rough sizing

Why this shape of host

FAQ

Comments

Leave a comment