Migrate from Bifrost to DVARA.
Move the operating controls with the traffic.

Move one workload in six steps

Use a workload that represents production, including its model naming, streaming mode, tools, structured output and failure cases.

1. 1

   Write down the current contract

   Record the Bifrost host, model names, providers, virtual-key headers, routing rules, budgets, rate limits, plugins and metadata in use. Save representative successful and failed responses.

2. 2

   Choose the DVARA operating model

   DVARA Open Source provides a file-configured LLM data plane. The Enterprise platform adds Flightdeck, central persistence, workspaces and managed governance workflows. Test the deployment shape you plan to operate.

3. 3

   Rebuild access, routes and controls

   Add provider credentials, create new DVARA keys, map only the public model names clients need, and rebuild routes, limits and governance. Bifrost and DVARA both allow unauthenticated inference by default; set DVARA_LLM_GATEWAY_REQUIRE_API_KEY=true when your DVARA deployment must require a key.

4. 4

   Send one complete governed request

   Replace the host, keys and model names with values from your test deployments. Bifrost commonly uses a provider-prefixed model and virtual key; the DVARA route decides the provider behind the public model name.

   Before — Bifrost

   curl https://<your-bifrost-host>/v1/chat/completions \
     -H 'Authorization: Bearer <bifrost-virtual-key>' \
     -H 'Content-Type: application/json' \
     -d '{
       "model": "openai/gpt-4o-mini",
       "messages": [{
         "role": "user",
         "content": "Summarize the refund policy."
       }],
       "temperature": 0
     }'

   After — DVARA

   curl https://<your-dvara-host>/v1/chat/completions \
     -H 'Authorization: Bearer <your-dvara-key>' \
     -H 'Content-Type: application/json' \
     -d '{
       "model": "gpt-4o-mini",
       "messages": [{
         "role": "user",
         "content": "Summarize the refund policy."
       }],
       "temperature": 0
     }'

   A representative successful response looks like this:

   {
     "id": "chatcmpl-01JQ8BIFROST2DVARA",
     "object": "chat.completion",
     "model": "gpt-4o-mini",
     "choices": [{
       "index": 0,
       "message": {
         "role": "assistant",
         "content": "Refunds are available within 30 days."
       },
       "finish_reason": "stop"
     }],
     "usage": {
       "prompt_tokens": 16,
       "completion_tokens": 9,
       "total_tokens": 25
     }
   }

   In the Enterprise platform, find the request under the expected workspace in Flightdeck. In Open Source, check the audit sink you configured. Confirm the provider, policy outcome, status and token usage.

5. 5

   Test the paths that expose differences

   Test streaming and non-streaming calls, tools, structured output, invalid keys, unknown models, limit boundaries, policy denials, provider errors and failover. Compare response shape, status, token accounting, latency and recorded evidence.

6. 6

   Cut over in stages

   Keep the Bifrost and DVARA URL/key pairs in runtime configuration. Move one workload, watch agreed error, latency, cost and denial thresholds, then expand only after the acceptance window passes.