LiteLLM hosting on Selfhost.dev puts the LiteLLM gateway and its own PostgreSQL 16 on a dedicated server we provision, secure and run. Your apps call one OpenAI-compatible endpoint with scoped keys. The gateway forwards each call to the provider with your own API key, so nobody takes a cut.
Two containers on one server. Requests are never metered; providers bill you directly.
LiteLLM is an open source LLM gateway: one OpenAI-compatible endpoint in front of OpenAI, Anthropic, Gemini, Bedrock and many more. LiteLLM hosting is the server, database and HTTPS that gateway needs to run for your team.
Teams calling several model providers who want one endpoint, per-app keys with budgets and a record of what each app spends.
When your apps already use more than one provider or you need spend limits per team. If you would rather buy credits and get one bill, a hosted router fits better.
LiteLLM proxy guides still point at the 2024 liteLLM-proxy repository and python main.py. The proxy now ships as a container image. This template runs ghcr.io/berriai/litellm-database v1.98.0 beside PostgreSQL 16, which keeps your virtual keys, budgets, spend and model settings between restarts.
Your apps, agents and teammates
TLS and routing
LiteLLM v1.98.0
PostgreSQL 16
Master key
Your model providers
| Component | Image | What it does |
|---|---|---|
| LiteLLM | ghcr.io/berriai/litellm-database:v1.98.0 | The gateway and its admin dashboard on port 4000 |
| PostgreSQL | postgres:16-alpine | Virtual keys, budgets, spend and models, on major version 16 |
| Volume | pg-data | The database files |
An LLM gateway holds every provider key your company uses, so its master key matters. LiteLLM's older docs used sk-1234 and copied setups often keep it. Here the master key comes from your password and the dashboard login is set outright, both before LiteLLM first starts.
| Running LiteLLM by hand | LiteLLM on Selfhost.dev |
|---|---|
A master key left at sk-1234 from older docs | LITELLM_MASTER_KEY built from your password before the first boot |
| A dashboard login that falls back to defaults | UI_USERNAME and UI_PASSWORD set outright |
| Edit config.yaml and restart for each new model | Models and provider keys added in the dashboard, no redeploy |
Run Postgres and point DATABASE_URL at it | PostgreSQL 16 in its own container; migrations run when LiteLLM starts |
| Put HTTPS in front of port 4000 | A certificate per hostname, renewed for you |
LiteLLM hosting requirements are small because the gateway forwards calls instead of running models. One proxy worker and PostgreSQL 16 run on 4 GB, the smallest server we offer. Redis matters when several LiteLLM instances share rate limits; this template runs one worker, so it does without.
Two containers, one server, no config file to write.
Launch opens the LiteLLM template in the console. Pick a location and set the admin password; it also becomes your master key, prefixed sk-.
The template deploys in about six minutes, database migrations included. Then sign in at /ui as admin with that password.
In the dashboard, add a model with your provider's API key. Then create a scoped virtual key for each app, so no app holds the master key.
LiteLLM hosting here is the server and nothing else. The gateway is free, the template adds no fee and requests are never metered. A European 4 GB server runs it at $0.02/hr, about $15 to $16 a month, billed hourly against prepaid credits. Model usage goes on your own provider accounts.
| Item | Cost |
|---|---|
| LiteLLM and PostgreSQL | $0 |
| Requests, tokens and virtual keys | $0 at any volume |
| Bandwidth | $0, traffic is included in the server price |
| Optional daily backups | +20% of the server price |
| Model usage | Billed by your providers to your own accounts |
| Server | RAM, cores, disk | Where | Rate | About a month |
|---|---|---|---|---|
cx23 (recommended) | 4 GB, 2 cores, 40 GB | Europe | $0.02/hr | $15.18 |
cax11 (ARM) | 4 GB, 2 cores, 40 GB | Europe | $0.02/hr | $16.35 |
cpx21 | 4 GB, 3 cores, 80 GB | US | $0.12/hr | $87.75 |
cpx22 | 4 GB, 2 cores, 80 GB | Singapore | $0.10/hr | $72.56 |
cx33 (more room) | 8 GB, 4 cores, 80 GB | Europe | $0.03/hr | $23.36 |
Monthly figures assume 730 hours.
A server with a card on file or a past top-up is powered off and held for 24 hours; calls to the gateway fail until then. A top-up in that window restarts it. Welcome-credit-only servers are deleted at zero.
Price another size or region in the calculator, no signup.
LiteLLM pricing starts at zero: the open source gateway is free to self host, by LiteLLM's own pricing page. Enterprise adds governance and support, sized to your usage and never priced per token. A hosted router such as OpenRouter skips the server but adds a fee on the credits you buy.
| Option | What you pay | Who holds the provider keys |
|---|---|---|
| LiteLLM open source, self hosted | Nothing for the software | You |
| LiteLLM Enterprise | Sized to your usage, never per token; no public list price | You |
| OpenRouter, a hosted router | Model cost plus a 5.5% platform fee (8% on Business) | OpenRouter |
| LiteLLM on Selfhost.dev | $0.02/hr for 4 GB in Europe, plus your provider bills | You |
Figures from litellm.ai/pricing and openrouter.ai/pricing.
When your model spend is small or you want one bill across hundreds of models without opening provider accounts. A hosted router does that for its fee. Here you hold the keys, the budgets and the spend records.
The LiteLLM template deploys in about six minutes. That is the platform's own budget, not an estimate.
Nothing here in your way? The template deploys in about six minutes.
Launch LiteLLMLiteLLM backup and restore comes down to one database. Virtual keys, teams, budgets, spend and the models you added all live in PostgreSQL. The server can take an optional daily image of everything on it; for a portable pg_dump, ask us for container access over SSH.
| Method | What it covers | Where it lives |
|---|---|---|
| Daily server backup (optional, +20%) | LiteLLM, PostgreSQL and the config; the last 7 kept | With the server; deleted with it |
pg_dump of the LiteLLM database | Keys, teams, budgets, spend and model settings | Your storage, via container access over SSH, granted on request |
Tell us how many apps call the gateway and roughly how many requests a day. We reply with a server size and its monthly cost.
Yes. The open source LiteLLM gateway is free to self host, forever, by LiteLLM's own pricing page. Enterprise adds governance, security and support, sized to your usage and never priced per token. Here you pay for the server by the hour and your model providers directly. See pricing.
For virtual keys, budgets and spend tracking, yes. LiteLLM can run from a config file alone, but key generation and spend limits need a database behind it. This template runs PostgreSQL 16 beside the gateway and applies LiteLLM's migrations each time it starts.
Point any OpenAI-compatible SDK at your LiteLLM address and use a virtual key from the dashboard as the API key. Ask for a model by the name you gave it in LiteLLM; the gateway calls the provider with your stored key and records the spend against that key.
Yes. Our MCP server can create the project, deploy the LiteLLM template and manage its domain from Claude, Cursor or any MCP client. Adding providers and minting virtual keys then happens in LiteLLM's own dashboard, signed in as admin with your password.
Selfhost.dev is a managed self hosting platform for developers and small teams. Open source apps and databases on dedicated servers you control, billed by the hour, in six locations across three continents.
Best for: LiteLLM, AnythingLLM, n8n and managed PostgreSQL, MySQL, Redis, ClickHouse, OpenSearch and Kafka.
