litestar-settings

Typed settings for Litestar apps, built on msgspec.Struct - like pydantic-settings, but msgspec instead of Pydantic.

Features

• Framework-agnostic core. litestar_settings.core has no dependency on Litestar; a settings schema is just a msgspec.Struct.
• Composable sources. EnvSource, DotEnvSource, TomlSource, YamlSource, SecretsDirSource, MappingSource - each reads from one place and returns a raw nested dict.
• Explicit merge order. You pass load_settings(cls, sources) an ordered list; later sources override earlier ones, recursively for nested dicts. No hidden "env beats file beats default" precedence.
• One validation pass. All sources are merged into a single raw mapping first; that mapping is converted into your Struct exactly once via msgspec.convert().
• Secret[T] fields. Wrap secret-bearing fields (api_key: Secret[str]) so they never leak into repr()/logs by accident - built automatically, no dec_hook wiring needed.
• Thin Litestar plugin. SettingsPlugin builds the Struct once, exposes it through DI, and fails app boot - not the first request - if the merged config doesn't satisfy the schema.
• No hot-reload. Settings load once, at startup - see Why no hot-reload.

Install

pip install litestar-settings              # msgspec + python-dotenv, incl. DotEnvSource
pip install 'litestar-settings[yaml]'      # adds YamlSource (ruamel.yaml)
pip install 'litestar-settings[litestar]'  # adds the Litestar plugin's dependency

TOML support needs no extra - it uses the stdlib tomllib (Python 3.11+).

Quick start

import msgspec
from litestar_settings import DotEnvSource, EnvSource, MappingSource, TomlSource, load_settings


class DBSettings(msgspec.Struct):
    host: str
    port: int = 5432


class Settings(msgspec.Struct, kw_only=True):
    debug: bool = False
    db: DBSettings


settings = load_settings(
    Settings,
    [
        TomlSource("settings.toml"),          # 1. defaults / base config
        DotEnvSource(".env", prefix="APP_"),  # 2. local dev overrides
        EnvSource(prefix="APP_"),             # 3. real environment wins
        MappingSource({"debug": True}),       # 4. explicit override, e.g. from a CLI flag
    ],
)

Sources are applied left to right; each later source deep-merges onto the result of the ones before it. Nested keys work through the same __ convention across EnvSource, DotEnvSource, and SecretsDirSource:

APP_DEBUG=true
APP_DB__HOST=localhost
APP_DB__PORT=6543

becomes {"debug": "true", "db": {"host": "localhost", "port": "6543"}}, which load_settings then converts (lax by default, so "6543" becomes 6543: int).

Pass strict=True to load_settings to disable that coercion and require exact types.

Usage

Secrets files (Docker/Kubernetes style)

from litestar_settings import SecretsDirSource

# /run/secrets/db__password -> {"db": {"password": "<file contents>"}}
SecretsDirSource("/run/secrets")

Secret fields

msgspec.Struct's default repr() prints every field verbatim - including secrets. Wrap secret-bearing fields in Secret[T] so they don't end up in logs or tracebacks by accident:

import msgspec
from litestar_settings import Secret, load_settings, TomlSource


class Settings(msgspec.Struct, kw_only=True):
    api_key: Secret[str]
    db_port: Secret[int] = Secret(5432)


settings = load_settings(Settings, [TomlSource("settings.toml")])

print(settings)                       # Settings(api_key=Secret('**********'), db_port=Secret('**********'))
settings.api_key.get_secret_value()   # the real value

load_settings() builds Secret[...] fields automatically - the wrapped type (str, int, ...) still goes through the normal msgspec.convert() coercion, so Secret[int] from an env var string works the same as a plain int field would. No dec_hook wiring needed; pass your own dec_hook for other custom types and it's consulted after Secret.

Secret only guards repr()/str() - it doesn't stop you from serializing the wrapped value elsewhere (e.g. msgspec.json.encode(settings.api_key.get_secret_value())), and encoding a Secret-typed field back to JSON needs its own enc_hook since msgspec doesn't know the type.

Custom sources

Source is a structural Protocol - any object with a zero-arg load() method returning a raw nested mapping satisfies it, no subclassing required:

from litestar_settings import RawMapping


class RedisSource:
    def __init__(self, client: Redis, key: str) -> None:
        self._client = client
        self._key = key

    def load(self) -> RawMapping:
        return json.loads(self._client.get(self._key) or "{}")

RawMapping is just dict[str, object], exported for readability - using the plain dict[str, object] annotation works identically.

Litestar usage

from __future__ import annotations

import msgspec
from litestar import Litestar, get
from litestar.config.app import AppConfig
from litestar.di import NamedDependency
from litestar_settings import EnvSource, TomlSource
from litestar_settings.plugin import SettingsPlugin


class Settings(msgspec.Struct, kw_only=True):
    debug: bool = False
    api_key: str


def apply_app_config(settings: Settings, app_config: AppConfig) -> AppConfig:
    app_config.debug = settings.debug
    return app_config


settings_plugin = SettingsPlugin(
    Settings,
    [TomlSource("settings.toml"), EnvSource(prefix="APP_")],
    apply_app_config=apply_app_config,  # optional - see below
)


@get("/status")
async def status(settings: NamedDependency[Settings]) -> dict:
    return {"debug": settings.debug}


app = Litestar(route_handlers=[status], plugins=[settings_plugin])

Fail-fast timing

By default SettingsPlugin builds Settings lazily behind a cached factory and forces that first build from an on_startup hook - so a missing api_key fails ASGI lifespan startup (litestar run, AsyncTestClient(app=app), ...), not the first request that happens to depend on settings. Constructing Litestar(...) itself never fails on bad config; importing the app module stays cheap even if the environment isn't fully set up yet (useful for litestar routes, schema generation, etc.).

Pass apply_app_config when you want native AppConfig fields (debug, openapi_config, ...) derived from Settings. Since AppConfig must be finalized synchronously inside on_app_init, this switches the plugin to build Settings eagerly at that point instead of at on_startup - still fail-fast, just earlier (at Litestar(...) construction time).

Custom dependency key

SettingsPlugin(Settings, [...], dependency_key="app_settings")

@get("/status")
async def status(app_settings: NamedDependency[Settings]) -> dict: ...

Design

Core is framework-agnostic. litestar_settings.core has no dependency on Litestar and no magic layered on top of typing. A settings schema is just a msgspec.Struct. Its only required dependencies are msgspec and python-dotenv (for DotEnvSource); YamlSource and the Litestar plugin are behind extras - see Install.

Sources are plain objects with a load() method. EnvSource, DotEnvSource, TomlSource, YamlSource, SecretsDirSource, MappingSource - each reads from one place and returns a raw nested dict. None of them validate or coerce anything.

Merge order is an explicit list, not hidden precedence. You pass load_settings(cls, sources) an ordered list; later sources override earlier ones, recursively for nested dicts. There is no built-in "env beats file beats default" rule baked into the source types themselves.

One convert() pass at the end. All sources are merged into a single raw mapping first; that mapping is converted into your Struct exactly once via msgspec.convert(). Sources never touch msgspec at all.

Secret[T] is built into that one pass. load_settings() composes a dec_hook that recognizes Secret[...] field types and wraps them automatically; a dec_hook you pass yourself is consulted for anything Secret doesn't handle. See Secret fields.

Litestar layer is a thin wrapper on top. SettingsPlugin builds the Struct once, exposes it through DI (Provide(get_settings, use_cache=True)), and fails app boot - not the first request - if the merged config doesn't satisfy the schema.

Load-once at startup, no hot-reload. See Why no hot-reload.

Why no hot-reload

Settings load once, at startup. There's no file watcher, no signal handler, no live-reload path.

• Predictability. If config could change under a running process, two reads of settings.db_url within the same request could theoretically disagree. That's a class of bug worth not having.
• Fail-fast stays simple. With validation only at startup, the service either boots with a valid config or doesn't boot. Hot-reload forces a decision about what to do with an invalid config arriving mid-flight (roll back? crash? ignore the bad value?) - a separate problem this library doesn't try to solve.
• Most settings don't need it. DB URLs, secret keys, uptime-scoped feature flags - a restart (systemd restart, a new container) is cheap and makes the reload story trivial.

If you need live-updated values (e.g. a feature flag toggled without a restart), that's a different problem - a small polling/pubsub-backed accessor - and deliberately out of scope here.

Development

uv sync --all-extras

uv run ruff check
uv run ruff format --check
uv run pyrefly check
uv run pyrefly coverage check

uv run pytest

License

BSD 3-Clause. See LICENSE.