Must have
Installation requirements
- PostgreSQL is required for virtual keys, teams, users and spend tracking.
Application hosting
Hosting LiteLLM Proxy
LiteLLM Proxy is an AI gateway that puts one OpenAI-compatible API in front of many model providers, with virtual keys, budgets and spend tracking. It routes requests to models hosted elsewhere, so it needs CPU and memory, not a GPU.
Requirements last checked against the project's documentation.
A sensible first server for LiteLLM Proxy is 2 vCPU, 8 GB RAM and 20 GB of storage. The cheapest listed plan that fits is OVHcloud VPS-2 (US) at $8.50/month.
Understand the setup
Find the software and services needed to run the application.
See requirementsBudget your resources
Plan memory, storage and CPU for the way you will use it.
Open resource plannerBuild your shortlist
Compare plans for the starting size, then review the provider and billing terms.
See plansRecommended
Planning guidance
- Add Redis once you run more than one proxy instance; it holds rate limits, router state and shared cache.
- Pin a release tag of the official image instead of latest so upgrades and rollbacks are predictable.
Choosing a server for LiteLLM Proxy
Size by worker count: LiteLLM gives each worker 1 vCPU and 4 GiB of memory. Model inference happens at the providers it routes to, outside this server budget.
Where to start
A sensible first server for LiteLLM Proxy
- vCPU
- 2
- RAM
- 8 GB
- Disk
- 20 GB
LiteLLM's production guide sizes 1 vCPU and 4 GiB per worker and, on a single VM, one worker per vCPU, so a 2 vCPU server runs two workers in 8 GB. Running PostgreSQL on the same server and the disk size are our estimates.
LiteLLM production best practices · Size it for your workload
Our shortlist
Cheapest plans that meet this size
-
$8.50/mo list price
-
$14.99/mo list price
-
$20.00/mo list price
-
$21.99/mo list price
-
$26.99/mo list price
-
$32.12/mo list price
-
$32.88/mo list price
-
$36.59/mo list price
One cheapest plan per provider, based on each provider's standard list price. Check the plan page for introductory and renewal pricing before you buy.
Deploy with
Supported paths
The host must provide
Runtime
- PostgreSQL
- Required for keys, teams, users and spend logs
- Redis
- Required when more than one proxy instance runs
- Model providers
- API keys for the hosted or self-hosted models the proxy routes to
Storage
Persistent data
The proxy itself is stateless. Keys, teams, spend logs and models added in the Admin UI live in PostgreSQL, so size and back up the database; spend logs grow with request volume.
Step by step
Plan a LiteLLM Proxy deployment
-
Set the master and salt keys
Set LITELLM_MASTER_KEY (it must start with sk-) and a long random LITELLM_SALT_KEY in a secret store. The salt key encrypts stored provider credentials and cannot be changed after you add models.
Official instructions -
Connect PostgreSQL and size the workers
Point DATABASE_URL at PostgreSQL, give each worker 1 vCPU and 4 GiB, and cap the database connection pool so all workers together stay within what PostgreSQL accepts.
Official instructions -
Add Redis before scaling out
Run a single instance until you need more; a second instance needs Redis for rate limits and router state.
Official instructions
Security
Before you go public
- Keep the master and salt keys in a secret manager and never commit them.
- Put the proxy and Admin UI behind HTTPS and set trusted proxy ranges so sign-in rate limits see real client addresses.
- Do not expose the PostgreSQL or Redis ports publicly.
Cost
Beyond the server price
- The server is small next to model usage: provider token charges usually dominate the bill.
- Each extra worker adds 1 vCPU and 4 GiB, and a managed PostgreSQL or Redis instance adds its own monthly cost.
Want the full monthly bill? Use the hosting cost calculator. Add storage, backups and other costs to your estimate.