The fundamental value of LiteLLM is abstraction. Instead of writing separate integration code for OpenAI, Anthropic, Google VertexAI, AWS Bedrock, Azure, Cohere, HuggingFace, and dozens of other providers, developers write one completion call and swap the model string. LiteLLM handles the translation between provider-specific API formats, authentication methods, and response schemas automatically.

The proxy server component turns LiteLLM into a production-grade AI gateway. Teams deploy it as a centralized endpoint that all their applications call. It provides virtual API keys with per-key budget limits and rate limiting, automatic load balancing across multiple model deployments, cost tracking and spend analytics per key/user/team, guardrails and content filtering, and detailed request/response logging. This makes it practical for organizations running multiple AI-powered applications that need centralized control over LLM usage and spending.

A skill wrapping LiteLLM gives an AI agent the ability to route requests to the optimal model for each task. The agent could use cheaper models for simple tasks and premium models for complex reasoning, all through the same interface. The proxy supports /chat/completions, /embeddings, /images, /audio, /batches, /rerank, and the new /a2a agent-to-agent protocol.

Installation requires just pip install litellm for the SDK or pip install litellm[proxy] for the gateway server. LiteLLM is MIT-licensed with active daily releases and comprehensive documentation at docs.litellm.ai.