# LLM provider specific environment variables. Check out the following resource # to learn more about all the providers (and, consequently, environment # variables) LiteLLM supports: https://docs.litellm.ai/docs/providers OPENAI_API_KEY= # OPTIONAL: Set the Anthropic API key if you still want to use Anthropic models # (see the explanation to the REMAP_* variables below). #ANTHROPIC_API_KEY= # OPTIONAL: Override base URLs if needed #OPENAI_BASE_URL=https://api.openai.com/v1 #ANTHROPIC_BASE_URL=https://api.anthropic.com # OPTIONAL: Authentication for the LiteLLM server (recommended when exposing # the server beyond your machine). Set a strong random key to require clients # to authenticate to the LiteLLM server. # # When set, clients must present this key as their API key when calling the # proxy. For Claude Code CLI, you can pass it inline like so: # ``` # ANTHROPIC_API_KEY="" \ # ANTHROPIC_BASE_URL=http://localhost:4000 \ # claude # ``` # (If you've previously signed in with Claude Code CLI, run `claude /logout`.) #LITELLM_MASTER_KEY= # OPTIONAL: Change model remaps if you need to (the values you see below are # the defaults). # # NOTE: If you don't want to remap one or more Claude models at all and want to # keep using the original models by Anthropic, uncomment the respective # variables, but delete their values (leave nothing to the right of the equal # sign). Alternatively, you can set them to map back to the Anthropic models of # your choosing explicitly. #REMAP_CLAUDE_HAIKU_TO=gpt-5.1-codex-mini-reason-none #REMAP_CLAUDE_SONNET_TO=gpt-5-codex-reason-medium #REMAP_CLAUDE_OPUS_TO=gpt-5.1-reason-high # OPTIONAL: You can turn off the prompt injection that forces non-Claude models # to use only one tool at a time. # # ATTENTION: Turning it off is NOT recommended. GPT-5, when used with medium or # high reasoning effort and without such injection, attempts to make multiple # tool calls at once quite often, and that causes Claude Code CLI to silently # stop processing the request (the CLI does not support multiple tool calls in # a single response). #ENFORCE_ONE_TOOL_CALL_PER_RESPONSE=false # OPTIONAL: Whether to convert ChatCompletions API requests to Responses API # format for ALL non-Claude models (true), or only for the OpenAI models that # don't support ChatCompletions API (false or unset, RECOMMENDED). # # The OpenAI models that don't support ChatCompletions API are: GPT-5-Codex, # GPT-5.1 Codex, GPT-5 Pro and few others (see https://platform.openai.com/docs/models) #ALWAYS_USE_RESPONSES_API=true # OPTIONAL: Langfuse configuration for logging LiteLLM request/response traces. # Useful for debugging. # # NOTE: If you set these keys, make sure to install Langfuse with either # `uv sync --all-extras` or `uv sync --extra langfuse` #LANGFUSE_SECRET_KEY="sk-..." #LANGFUSE_PUBLIC_KEY="pk-..." #LANGFUSE_HOST="https://cloud.langfuse.com" # OPTIONAL: Alternative logging of LiteLLM request/response traces (as well as # potential conversions between ChatCompletions API and Responses API) in the # form of local markdown files written to `.traces/` folder. Makes it easier to # feed the traces into AI Coding Assistants to fix things. #WRITE_TRACES_TO_FILES=true PYTHONUNBUFFERED=1