Skip to main content

​
Overview

Firebender can integrate with your LiteLLM deployment in two ways:
  • Model discovery: Firebender reads available models from LiteLLM
  • Request routing: Firebender sends chat completions through LiteLLM for models marked with provider: 'lite-llm'

​
1. Expose model discovery endpoints

Firebender checks LiteLLM model metadata using:
  • /model/info
  • /v1/models as a fallback
Make sure at least one of those endpoints is reachable from Firebender.

​
2. Create a discovery/admin key

Use a LiteLLM key that can read model metadata. Firebender uses this key to discover which models your deployment exposes.

​
3. Create virtual keys for request auth

Firebender requires a LiteLLM virtual key for actual request execution. That means:
  • the discovery/admin key alone is not enough to send requests
  • users who route through LiteLLM need a valid virtual key configured in Firebender

​
Example LiteLLM alias

A common pattern is to expose a friendly alias from LiteLLM and map it to a real upstream provider model. 
model_list:
  - model_name: bedrock-claude-sonnet-4-6
    litellm_params:
      model: bedrock/anthropic.claude-sonnet-4-20250514-v1:0
With a setup like this:
  • Firebender can show bedrock-claude-sonnet-4-6 in its merged model list
  • LiteLLM still forwards requests to the real upstream Bedrock Anthropic model

​
What Firebender expects from LiteLLM metadata

Firebender uses LiteLLM model metadata to build merged model definitions, including fields like:
  • model alias / ID
  • reasoning support
  • vision support
  • PDF input support
  • token limits
  • pricing metadata when available

​
Firebender-side configuration

For the Firebender steps, see Firebender on LiteLLM.

​
Troubleshooting

​
/model/info is unavailable

That is okay as long as /v1/models works. Firebender falls back automatically.

​
Firebender shows the model, but requests fail with a virtual key alert

That means discovery is working, but request auth is not. Add a LiteLLM virtual key in Firebender for the user who is sending requests.

​
Aliases work in LiteLLM but not in Firebender

Double-check the alias exposed by LiteLLM and make sure it is stable and points to the upstream model you expect.