LiteLLM Free code examples in Python and JavaScript

The proxy extra, and a Postgres instance. The database is not optional if you intend to enforce anything.

pip install 'litellm[proxy]'

docker run -d -p 5432:5432 -e POSTGRES_PASSWORD=dev postgres:17

export LITELLM_MASTER_KEY="sk-master-dev"    # generate a long random one
export DATABASE_URL="postgresql://postgres:dev@localhost:5432/postgres"

model_name is the alias callers ask for; litellm_params says what it really is. That indirection is the point — swapping a model for every caller is one line here and no change anywhere else.

# config.yaml
model_list:
  - model_name: fast                       # what callers request
    litellm_params:
      model: openai/gpt-5.4-mini           # what it actually is
      api_key: os.environ/OPENAI_API_KEY

  - model_name: smart
    litellm_params:
      model: anthropic/claude-sonnet-4.6
      api_key: os.environ/ANTHROPIC_API_KEY

general_settings:
  master_key: os.environ/LITELLM_MASTER_KEY
  database_url: os.environ/DATABASE_URL    # REQUIRED for budgets

litellm_settings:
  max_budget: 500                          # proxy-wide ceiling
  budget_duration: 30d

Provider keys stay in the proxy's environment. Everyone else gets a LiteLLM key with a budget and an allow-list — and revoking one is an API call rather than a rotation across every repository that had it.

litellm config.yaml

# A team with a monthly ceiling.
curl -sX POST 'http://localhost:4000/team/new' \
  -H "Authorization: Bearer $LITELLM_MASTER_KEY" \
  -H 'Content-Type: application/json' \
  -d '{
        "team_alias": "search-squad",
        "max_budget": 100,
        "budget_duration": "30d",
        "rpm_limit": 99
      }'

# A key inside that team, with its own smaller budget and a
# restricted model list.
curl -sX POST 'http://localhost:4000/key/generate' \
  -H "Authorization: Bearer $LITELLM_MASTER_KEY" \
  -H 'Content-Type: application/json' \
  -d '{
        "team_id": "<team-id-from-above>",
        "max_budget": 10,
        "budget_duration": "30d",
        "models": ["fast"]
      }'

An OpenAI client against the proxy. The application never sees a provider key, and the model name is the alias from the config rather than a vendor's id.

import os
from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:4000",
    api_key=os.environ["LITELLM_VIRTUAL_KEY"],   # not the provider key
)

completion = client.chat.completions.create(
    model="fast",             # the alias, not openai/gpt-5.4-mini
    messages=[{"role": "user", "content": "Explain a database index."}],
)

print(completion.choices[0].message.content)

The proxy returns 400 with a budget message — not 429, and not a silent downgrade. Handle it explicitly, because a retry loop written for rate limits will hammer a spent budget forever.

import openai

try:
    completion = client.chat.completions.create(
        model="fast",
        messages=[{"role": "user", "content": "Explain a database index."}],
    )
except openai.BadRequestError as error:
    message = str(error)

    # Budget exhaustion arrives as a 400, so generic 429 retry logic
    # will not catch it and generic 400 handling will call it a bug.
    if "budget" in message.lower():
        raise RuntimeError("Team budget exhausted — raise it or wait.") from error

    raise

# Check spend before it bites:
#   curl http://localhost:4000/key/info \
#     -H "Authorization: Bearer $LITELLM_VIRTUAL_KEY"