Best VPS for CrewAI (2026): Multi-Agent Hosting Without the Bloat
CrewAI gets recommended for everything from research agents to customer support bots, and the spec recommendations all over Reddit are wildly inconsistent. Some threads say 32 GB RAM, others say 1 GB. I have run CrewAI in production for four months across three different VPS configs. Here is what actually fits.
The Real CrewAI Footprint
The CrewAI library is genuinely small. A Python process with one crew loaded sits at around 200 to 300 MB. Adding more agents to the same crew adds maybe 50 MB each. The number of agents is not the cost driver.
What is the cost driver:
- LLM client buffers (especially with streaming and structured output)
- The memory feature, which loads embeddings models if enabled
- Tool overhead, especially browser tools or local code execution
- Concurrent crew runs, since each run holds open multiple HTTP connections to the model API
For a typical 4 to 6 agent crew with web search and a couple of API tools, 4 GB RAM and 2 vCPU handles it comfortably. Anyone telling you to provision 16 GB is sizing for self-hosted LLMs, not for CrewAI itself.
VPS Comparison for CrewAI
| Provider | Plan | vCPU | RAM | Disk | Monthly | Best fit |
|---|---|---|---|---|---|---|
| Hetzner Cloud | CCX13 | 2 | 8 GB | 80 GB NVMe | 14.86 EUR | Production crews, EU |
| Contabo VPS | VPS S | 4 | 8 GB | 100 GB NVMe | 4.50 EUR | Hobby and side projects |
| DigitalOcean | Premium AMD 4 GB | 2 | 4 GB | 80 GB NVMe | 28 USD | US, low ops overhead |
| Hetzner CPX11 | CPX11 | 2 | 2 GB | 40 GB SSD | 5.18 EUR | Solo dev, single crew |
Hetzner Cloud CCX13: The reliable production pick
Dedicated CPU matters for CrewAI more than people realize, because the crew often spawns multiple parallel agent calls and shared CPU plans can introduce latency variance. The CCX13’s two dedicated cores give predictable performance even when 3 to 4 crew runs overlap.
Why I pick this for client work: predictable response times when multiple crews fire at the same time. Shared-CPU plans can stretch a 5-second crew run to 12 seconds during noisy-neighbor periods.
Pros worth knowing:
- Dedicated CPU keeps latency variance low
- NVMe handles the structured output logs without complaint
- Hetzner network has consistent sub-100 ms latency to all major LLM APIs
The trade-off: 14.86 EUR a month is roughly 3x the cheapest option. Worth it for production, overkill for experiments.
Get Hetzner: Hetzner Cloud.
Contabo VPS S: For experiments and side projects
4.50 EUR a month for 4 vCPU and 8 GB RAM is hard to argue with for non-critical workloads. CrewAI runs on it fine. The shared CPU shows up as occasional 300 to 500 ms latency stretches under load, which matters more for interactive use than batch jobs.
Pros:
- Cheapest realistic option for CrewAI
- 8 GB RAM leaves room for the memory feature plus larger embedding models
- 100 GB NVMe is plenty for crew state and logs
Real negative: provisioning takes 2 to 4 hours, which is painful if you iterate on infrastructure.
Try Contabo: Contabo VPS.
DigitalOcean Premium AMD: For US-based ops
If your team, model APIs, and integrations are all US-east, the US latency from DigitalOcean NYC3 makes a noticeable difference on crews that chain multiple tool calls. 28 USD a month is not cheap against Hetzner, but the snapshot system and managed databases reduce operational work.
Honest take: the value is in the platform, not the raw compute. If you do not care about the DigitalOcean ecosystem, Hetzner gets you more performance per dollar.
Get DigitalOcean: DigitalOcean.
Hetzner CPX11: The cheapest path that works
For a solo developer running a single crew interactively, the CPX11 with 2 GB RAM is enough. CrewAI fits, the LLM client buffers fit, you just cannot enable the memory feature with a large embedding model.
Use this for personal projects and demos. Move up the moment the work matters.
What I Would Pick
For client work or production CrewAI deployments: Hetzner CCX13. For experiments and hobby projects: Contabo VPS S. The CrewAI enterprise cloud is fine for low-volume use, but self-hosting on a properly sized VPS gives you more flexibility and costs less past 40 EUR a month of usage.
The full VPS picture sits at the SelfHostVPS comparison. CrewAI is stable enough that hosting requirements have not shifted much in 2026, the recommendations here should hold for a while.
Frequently asked questions
What VPS spec does CrewAI need for a 5-agent crew?
CrewAI itself is light, the Python process sits at around 200 to 400 MB resident memory regardless of crew size. A 5-agent crew with default tools fits inside 2 GB RAM. The memory pressure comes from the LLM client buffers and any embeddings if you use the memory feature. Plan for 4 GB to leave headroom, but you do not need 16 GB unless you also self-host the model.
Does CrewAI need a GPU on the same VPS?
No, in 90% of setups. CrewAI calls external LLM endpoints (OpenAI, Anthropic, Groq, Together) and the VPS only needs CPU. GPU becomes relevant only if you self-host the language model with Ollama or vLLM on the same machine. Most production CrewAI deployments use external models and run on CPU-only nodes.
How does CrewAI compare to LangGraph for hosting requirements?
CrewAI is lighter in steady state, the abstraction is thinner and the memory footprint smaller. LangGraph adds the LangChain dependency tree plus the graph state machine, which roughly doubles the resident memory for equivalent workflows. If hosting cost matters and your use case fits role-based collaboration, CrewAI wins on price.
Should I use CrewAI's enterprise cloud or self-host?
Self-host if you need data residency, want to integrate private tools, or expect to run more than 50 crew executions a day. The enterprise cloud is convenient for low-volume use and removes operational overhead. Break-even against a Hetzner CCX13 sits around 40 to 60 EUR a month of cloud usage depending on the plan you would pick.