Overview: LiteLLM provides a common interface for many model providers and can also run as a centralized AI gateway. It reduces provider-specific code and normalizes common request and error patterns.
What beginners can do: Beginners can start with one Python completion call, then switch the model identifier to compare providers. Teams can later introduce the proxy for virtual keys, budgets, routing, guardrails, and observability.
Recommended start: Create an isolated Python environment, install the current SDK, configure one provider key through an environment variable, and make one low-cost request. Do not deploy the proxy until the direct SDK path is understood.
Requirements and cost: At least one supported model provider or local compatible endpoint is needed. The gateway path adds a service, configuration store, authentication, and operational responsibilities.
Who it suits: Useful for developers who expect to compare or switch providers and teams that need centralized model access control.
Limits and safety: A unified interface cannot erase provider differences in model behavior, supported parameters, quotas, or pricing. Some capabilities and licensing arrangements differ between open and enterprise offerings. Keep real provider keys only on the server, issue scoped virtual keys to clients, redact request logs, set budgets and rate limits, and never expose an unauthenticated proxy to the internet.
Project and license: Repository content is provided under Repository license files; enterprise features differ; check the current LICENSE before reuse. Repository: https://github.com/BerriAI/litellm