The idea
The same code runs in several environments with different settings: database URLs, model names, log levels, feature flags, rate limits and API keys. Configuration comes from the environment, not from code (the twelve-factor principle), so one image runs everywhere.
Load and validate config at startup with a typed settings class, fail fast on missing values, and never default production secrets. Use different provider keys and budgets per environment so a staging bug cannot burn the production quota.
Keep non-secret config in version control (per-environment files) and secrets in a secret manager.
Config mistakes cause many outages: staging pointing at the production database, debug logging of prompts in prod.
Words to know
- Twelve-factor appA set of principles for deployable services, including config in the environment.
Try it without code
List your settings
- List every setting your app reads.
- Mark secret or not, and the value per environment.
Use it in code
Typed settings with pydantic-settings.
# pip install pydantic-settings
from typing import Literal
from pydantic_settings import BaseSettings, SettingsConfigDict
class Settings(BaseSettings):
model_config = SettingsConfigDict(env_file=".env", extra="ignore")
env: Literal["dev", "staging", "prod"] = "dev"
database_url: str
anthropic_api_key: str
chat_model: str = "claude-sonnet-5-5"
log_prompts: bool = False
daily_token_budget: int = 50_000
settings = Settings() # raises a clear error if a required value is missing
if settings.env == "prod":
assert not settings.log_prompts, "never log prompts in prod"Replace scattered os.environ calls in your app with one Settings class and separate dev and prod values.
Common mistakes
- Hard-coded model names in many files.
- Defaults that silently point at production.