LiteLLM

Multi-agent orchestration

LiteLLM exposes a single interface in front of every model, local or remote, making them interchangeable. It handles routing, aliases, virtual keys per user and team, quotas, cost tracking and logging.

Key points: Unified gateway · Routing and fallback · Virtual keys · Budgets and quotas · Cost tracking · Observability · Guardrails · Redis cache.

Unified gateway

LiteLLM presents a single OpenAI-compatible interface in front of more than a hundred providers (official list, "100+ LLMs" as of 2 September 2026). The same client call reaches a local model served by vLLM or llama.cpp just as it would reach a remote OpenAI, Anthropic, Amazon Bedrock, Azure or OpenRouter API, without changing the application code.

Routing, fallback and aliases

The gateway distributes load across multiple instances of the same model and falls back to a backup model in case of error or saturation. It handles retries, delays and cooldown periods. A group alias brings several models together behind a single logical name, which makes it possible to switch provider without touching the clients.

Virtual keys, budgets and quotas

Each user or team receives a virtual key with an assigned budget, token and requests-per-minute limits, and a list of allowed models. Provider keys stay hidden behind the gateway. Exceeding the budget blocks the key.

Cost tracking

LiteLLM calculates the cost of each call and aggregates it by key, user, team and model. The consumption history feeds budget tracking and dashboards, with resets according to the chosen period.

Observability and logging

Every request and every response is logged and exported to Langfuse, Prometheus or custom callbacks. Tracking covers latency, throughput, errors and tool-call detail, which makes usage traceable and auditable.

Guardrails, cache and administration

Pre- and post-call plugins filter inputs and outputs, apply privacy rules and strip sensitive content. A Redis cache pools identical responses and reduces cost and latency. An administration interface and an API manage keys, budgets and models. The gateway is self-hosted and keeps secrets on sovereign infrastructure.

License and source code

LiteLLM is implemented in Python (its publisher BerriAI announces a Rust core) and its source is public. Its LICENSE file places it under the MIT licence, except for the enterprise/ directory, which is governed by a separate commercial licence (re-read on 2 September 2026); this is why GitHub returns no standard identifier. Check that the features you rely on stay outside the enterprise scope before any commercial deployment.

Repository: BerriAI/litellm. Licence and implementation language are re-read from this source by the QDNA fact base. The version number is deliberately not published here: it would go stale between two upstream releases without anything catching it.