Moving from LiteLLM Cloud to a DVARA gateway you run.
What transfers, what does not, and how to check.

How do you migrate off LiteLLM Cloud?

For a plain OpenAI-SDK path the client change is small. The work is standing up the gateway you are pointing at, and then proving the paths your application actually uses still behave the same.

1. 1

   Inventory what you rely on today

   Which endpoints, models and providers you call; whether you stream; whether you use tools or structured outputs; and which LiteLLM concepts are load-bearing — virtual keys, teams, budgets, routing rules, callbacks. None of that transfers, so the list is what tells you how big the migration really is.

   Capture a set of representative requests and their responses now. They are what you will compare against.

2. 2

   Stand up a DVARA deployment to test against

   This is the part that has no equivalent in your current setup. DVARA is software you operate — gateway and console in your own environment, against your own database. The public images boot with no licence key, enough to exercise the compatibility surface; the governance engines need a licence.

   Read the quickstartSee pricing
3. 3

   Add provider credentials, then repoint the SDK

   Register your provider keys in the console, then change the base URL and key in your client. Keep both in configuration rather than hardcoded — it is what makes step five reversible.

   Before — LiteLLM Cloud

   from openai import OpenAI
   
   client = OpenAI(
       base_url=os.environ["LLM_BASE_URL"],
       api_key=os.environ["LLM_API_KEY"],
   )

   After — your DVARA gateway

   # LLM_BASE_URL=http://localhost:8080/v1
   # LLM_API_KEY=gw_...
   #
   # Same code. Only the two environment
   # variables change — which is also how
   # you change them back.

   localhost:8080 is the gateway's documented port; substitute whatever host you deployed it to.

4. 4

   Replay your representative requests and compare

   Response shape, status codes, token counts, and the paths easiest to get wrong: streaming, tool calls, structured outputs, an unknown model, a missing key, and whatever your retry logic does when a provider errors.

   Differences are not necessarily faults — a provider that cannot honour strict structured output is reported rather than hidden. Decide which differences you accept.

5. 5

   Move a slice, then the rest — and keep the way back

   Cut over one workload before all of them. Rollback is setting the two environment variables back to what they were, which is only true while they stay configuration.

   Turn on the governance you need once traffic is steady. Doing it in the same change as the cutover means a policy denial and a compatibility problem look identical.