Litellm Proxy

LiteLLM Proxy — Unified LLM API Gateway

Overview

LiteLLM Proxy is a standalone OpenAI-compatible API server that routes requests to 100+ LLM providers (OpenAI, Anthropic, Bedrock, Azure, Ollama, Groq, etc.) with unified credentials, load balancing, rate limiting, cost tracking, and caching. Run it as a Docker container or Python process — teams point their existing OpenAI SDK code at it and swap models without code changes.

GitHub: https://github.com/BerriAI/litellm (15k+ stars)

When to Use

• Multi-provider LLM routing with a single API key for your team
• Cost tracking and budget enforcement across LLM providers
• A/B testing between models without changing application code
• Load balancing across multiple API keys or provider accounts
• Adding caching, logging, or guardrails in front of any LLM
• Migrating from OpenAI to another provider transparently

Installation

# Install LiteLLM
pip install litellm[proxy]

# Or via Docker
docker pull ghcr.io/berriai/litellm:main-latest

Key Patterns / Usage

Quick Start — Config File

# litellm_config.yaml
model_list:
  - model_name: gpt-4o
    litellm_params:
      model: openai/gpt-4o
      api_key: os.environ/OPENAI_API_KEY

  - model_name: claude-3-5-sonnet
    litellm_params:
      model: anthropic/claude-3-5-sonnet-20241022
      api_key: os.environ/ANTHROPIC_API_KEY

  - model_name: local-llama
    litellm_params:
      model: ollama/llama3.2
      api_base: http://localhost:11434

general_settings:
  master_key: sk-my-master-key
  database_url: os.environ/DATABASE_URL  # optional, for persistence

# Start the proxy
litellm --config litellm_config.yaml --port 4000

# Or Docker
docker run -v $(pwd)/litellm_config.yaml:/app/config.yaml \
  -e OPENAI_API_KEY=$OPENAI_API_KEY \
  -p 4000:4000 ghcr.io/berriai/litellm:main-latest \
  --config /app/config.yaml

Client Usage (OpenAI SDK)

from openai import OpenAI

# Point at LiteLLM proxy
client = OpenAI(
    base_url="http://localhost:4000",
    api_key="sk-my-master-key",
)

# Use any model name from your config
response = client.chat.completions.create(
    model="claude-3-5-sonnet",  # routes to Anthropic
    messages=[{"role": "user", "content": "Hello!"}],
)
print(response.choices[0].message.content)

Load Balancing + Fallbacks

model_list:
  - model_name: gpt-4o
    litellm_params:
      model: openai/gpt-4o
      api_key: os.environ/OPENAI_API_KEY_1

  - model_name: gpt-4o
    litellm_params:
      model: openai/gpt-4o
      api_key: os.environ/OPENAI_API_KEY_2  # second key = load balanced

router_settings:
  routing_strategy: least-busy
  fallbacks: [{"gpt-4o": ["claude-3-5-sonnet"]}]
  context_window_fallbacks: [{"gpt-4o": ["gpt-4o-mini"]}]
  num_retries: 3
  timeout: 30

Rate Limiting and Budgets

general_settings:
  master_key: sk-master

litellm_settings:
  max_budget: 100  # $100 total budget

# Per-key budgets via API
# POST /key/generate
# {"max_budget": 10, "models": ["gpt-4o"], "duration": "30d"}

import httpx

# Generate a virtual key with budget
resp = httpx.post(
    "http://localhost:4000/key/generate",
    headers={"Authorization": "Bearer sk-master"},
    json={"max_budget": 5.0, "models": ["gpt-4o-mini"], "duration": "7d"},
)
virtual_key = resp.json()["key"]
print(f"Team key: {virtual_key}")

Caching

litellm_settings:
  cache: true
  cache_params:
    type: redis
    host: localhost
    port: 6379
    ttl: 600  # seconds

  # Or in-memory
  cache_params:
    type: local

Logging and Observability

litellm_settings:
  success_callback: ["langfuse", "helicone"]
  failure_callback: ["sentry"]

  langfuse_public_key: os.environ/LANGFUSE_PUBLIC_KEY
  langfuse_secret_key: os.environ/LANGFUSE_SECRET_KEY

Guardrails

litellm_settings:
  guardrails:
    - guardrail_name: "pii-guard"
      litellm_params:
        guardrail: presidio
        mode: pre_call  # mask PII before sending to LLM

Common Pitfalls

• Master key required: always set master_key in config; without it anyone can hit your proxy
• Model name mismatch: the model_name in config is what clients use, not the provider's actual model name
• Environment variables: use os.environ/VAR_NAME syntax in YAML config (not $VAR_NAME)
• Docker networking: when Ollama runs on host, use host.docker.internal not localhost
• Cost tracking needs DB: spend tracking requires a PostgreSQL database_url in general_settings
• Streaming proxying: streaming works but ensure your client handles SSE properly

Related Skills

• openrouter-litellm — using LiteLLM Python library (not proxy) with OpenRouter
• ollama-integration — running local models behind the proxy
• llm-routing-and-fallback — routing and fallback strategies
• api-rate-limiting — rate limiting patterns
• llm-observability — observability for LLM applications

GitNexus Index

tool: litellm-proxy
category: llm-gateway
tier: self-hosted
interface: openai-compatible
platform: cross-platform
stars: 15000+

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: UltronCore
• Source: UltronCore/claude-skill-vault
• License: MIT

Install and usage instructions live in the source repository linked above.