Best VPS for LibreChat (2026): Real Specs for the Multi-User Chat Server
LibreChat became the de facto self hosted ChatGPT alternative for teams in 2026. It supports the big providers, MCP, Agents, Artifacts, Code Interpreter, and multi user auth. That feature density also means the resource curve is wider than a simple chat UI like Open WebUI.
I run a LibreChat instance for a five user team on Hetzner and tested four different VPS tiers to understand where the breakpoints really are.
What LibreChat Actually Needs
Four resource sinks to size for:
- The API server. Node, the agent runtime, the MCP layer. 300 to 500 MB resident at idle.
- MongoDB. The bigger driver. Conversations, agent definitions, file metadata, all in Mongo. Plan for 1 GB plus growth.
- Meilisearch. The text search index. Roughly the size of your message corpus. Wants to stay in memory.
- Code Interpreter. Optional but heavy when enabled. Each execution spawns a sandboxed container.
The thing that catches people: the RAG API service is a separate process with its own embeddings cache. Default config pulls in a sentence transformer model and runs it on CPU. Plan for another 1 GB resident and a slow first request.
VPS Comparison for LibreChat
| Provider | Plan | vCPU | RAM | Disk | Monthly | Best fit |
|---|---|---|---|---|---|---|
| Hetzner Cloud | CCX13 | 2 | 8 GB | 80 GB NVMe | 14.86 EUR | Solo to small team, default pick |
| Contabo VPS | VPS S | 4 | 8 GB | 100 GB NVMe | 4.50 EUR | Budget, solo or pair use |
| DigitalOcean | Premium AMD 4 GB | 2 | 8 GB | 100 GB NVMe | 28 USD | US team, low load |
| Hetzner Cloud | CCX23 | 4 | 16 GB | 160 GB NVMe | 29.74 EUR | Larger team with Code Interpreter |
Hetzner Cloud CCX13: The default
This is what I run mine on. The Mongo and Meilisearch combination behaves well on Hetzner’s NVMe, and Falkenstein latency to Anthropic and OpenAI is under 80 ms, which makes the agent loop feel snappy. 8 GB RAM is enough for a small team with file search and MCP enabled.
Pros that matter here:
- NVMe handles the Mongo write pattern well
- Snapshots make it easy to roll back a bad LibreChat upgrade
- API plus Terraform if you want infra as code
Real downside: 8 GB ceiling on CCX13. Once you add Code Interpreter with multiple concurrent users, you outgrow this and move to CCX23.
Get Hetzner: Hetzner Cloud.
Contabo VPS S: Budget pick
4 vCPU and 8 GB for 4.50 EUR. The newer NVMe Contabo plans hold up fine for LibreChat as long as you do not turn on every optional service. The compromises are familiar:
- Slow provisioning, hours not seconds
- Higher outbound latency to model APIs from European Contabo nodes
- Support reply times can stretch on the cheapest tiers
For a personal or two person use case it is hard to argue against the price.
Get Contabo: Contabo VPS.
DigitalOcean Premium AMD 4 GB: For US east teams
The 8 GB Premium AMD plan is enough for a small US team with light agent use. NYC3 is the cheapest path to low latency model calls if your team is on the east coast. Snapshots are useful when LibreChat ships breaking changes.
Honest negative: 28 USD a month for 8 GB is steep against Hetzner. You pay for the region and the snapshot system.
Get DigitalOcean: DigitalOcean.
Hetzner Cloud CCX23: For real teams
Once you enable Code Interpreter and have multiple users running agents at the same time, the 8 GB tier feels tight. CCX23 with 16 GB RAM and 4 dedicated vCPU gives you headroom for concurrent agent runs without slowing the chat experience.
Get Hetzner: Hetzner Cloud.
Common Setup Gotchas
Three things to plan for:
- MongoDB needs a real persistent volume. Do not put it on ephemeral storage. Losing the database means losing every conversation and agent definition.
- Meilisearch wants to stay warm. Restart the box and the first few searches are slow. Either pre warm or accept that morning queries take a beat.
- MCP servers add latency. Every MCP tool the agent has access to gets queried during planning. Trim what each user can reach so you do not pay for tools you do not use.
What I would actually pick
If you are starting today:
- EU, default daily use: Hetzner CCX13
- EU, solo or pair, lowest spend: Contabo VPS S
- US east team: DigitalOcean Premium AMD 8 GB
- Team with Code Interpreter: Hetzner CCX23
For the broader self hosted AI picture, see the SelfHostVPS comparison. LibreChat ships releases monthly, so I refresh this page when major resource patterns change.
Frequently asked questions
What is the minimum VPS for LibreChat?
2 vCPU and 4 GB RAM if you keep the stack small (Node API, MongoDB, Meilisearch, RAG-API) and serve under five users. The bulk of the RAM goes to MongoDB and the Meilisearch index, with the Node API itself staying under 400 MB resident. Add code interpreter and the resource curve jumps to 8 GB.
Does LibreChat need a GPU?
No, unless you also self-host the model and the code interpreter is doing heavy ML work. The default stack is CPU bound, with all model inference happening at whatever provider you wire in. A CPU only Hetzner CCX or Contabo NVMe VPS is the normal setup.
Can I run LibreChat on a 5 dollar Contabo box?
Yes for a single user with three or four conversations a day. The break point is when you turn on MCP servers, code interpreter, or the file search RAG pipeline. At that point you want 8 GB RAM minimum or the agent calls start taking visible seconds.
How much disk should I plan for LibreChat?
Plan for 80 GB NVMe at minimum. MongoDB grows with every conversation, the file upload feature stores assets on disk, and the Meilisearch index doubles your text storage. After three months with a small team I have seen the data volume pass 22 GB.