Settings¶

One module, resolved from the MLANGO_SETTINGS_MODULE environment variable — manage.py sets it for you. Every default lives in mlango.conf.global_settings; override only what differs.

An unknown setting is an error rather than a silent no-op:

>>> settings.METASTOR
AttributeError: 'METASTOR' is not a known mlango setting. Settings must be
uppercase and declared in mlango.conf.global_settings.

Core¶

Setting	Default	Purpose
`BASE_DIR`	cwd	Everything relative resolves from here
`SECRET_KEY`	`""`	Extra entropy for hashing. Keep it out of git
`DEBUG`	`False`	Verbose tracebacks and autoreload
`INSTALLED_APPS`	`[]`	Apps whose declarations are loaded at startup
`APP_MODULES`	`("datasets", "models", "agents", "evals", "admin", "signals")`	Modules autodiscovered in each app

Metastore¶

METASTORE = {
    "URL": "sqlite:///mlango.db",   # or postgresql://user@host/mlango
    "ECHO": False,                  # log every SQL statement
    "POOL_PRE_PING": True,          # non-SQLite only
}

A relative SQLite path resolves against BASE_DIR, so the database does not follow your shell's working directory around. SQLite runs in WAL mode, so the admin can read while a training loop writes.

Move to Postgres when more than one worker records runs.

Storage¶

STORAGE = {
    "BACKEND": "mlango.storage.local.LocalStorage",
    "ROOT": "artifacts",
}

Checkpoints, materialised datasets and run outputs go here. Artifacts are recorded in the metastore by storage-relative name, so a run on one machine resolves on another.

STORAGE = {
    "BACKEND": "mlango.storage.s3.S3Storage",
    "ROOT": "s3://my-bucket/mlango",
    # Optional — anything S3-compatible: MinIO, R2, B2.
    "ENDPOINT_URL": os.environ.get("S3_ENDPOINT_URL"),
    "REGION": "eu-west-1",
}

pip install "mlango[s3]"

Credentials are boto3's, so the usual environment variables, instance roles and profiles all work and mlango never handles a secret it does not have to. Point BACKEND at your own Storage subclass for anything else — see Training somewhere else.

Training¶

Setting	Default	Purpose
`TRAINERS`	sklearn and torch	`{name: dotted path}` available to `Meta.trainer`
`DEVICE`	`"auto"`	`"auto"` picks CUDA when available
`SEED`	`1337`	Seeds python, numpy and torch at the start of every run
`DEFAULT_CALLBACKS`	`[]`	Callbacks appended to every run
`PREDICTION_LOG`	off	What a served model records about its requests

TRAINERS = {"lightgbm": "myproject.trainers.LightGBMTrainer"}

DEFAULT_CALLBACKS = [
    "mlango.training.callbacks.ProgressBar",
]

Note

DEFAULT_CALLBACKS is purely additive — metric recording is built into the framework, so emptying the list never costs you run history.

PREDICTION_LOG = {
    "ENABLED": True,   # off by default
    "SAMPLE": 0.05,    # keep 5% of predictions
    "MAX_ROWS": 100_000,
}

The log is what manage.py drift measures against a version's training profile. It is off by default because it is a copy of user input in a database — a decision a project makes rather than one it wakes up with. See Monitoring.

Agents¶

Setting	Default	Purpose
`PROVIDERS`	anthropic and echo	`{name: dotted path}` available to `Meta.provider`
`DEFAULT_PROVIDER`	`"anthropic"`	Provider when an agent does not name one
`DEFAULT_AGENT_MODEL`	`"claude-opus-5"`	Model when an agent does not name one
`DEFAULT_THINKING`	`"adaptive"`	Thinking mode; `None` omits the parameter
`AGENT_MAX_STEPS`	`12`	Hard stop on the tool-use loop
`TRACING`	`True`	Record spans for every model and tool call

A scaffolded project starts on "echo" so it runs with no credentials. Switch to "anthropic" and export ANTHROPIC_API_KEY when you want a real model.

Admin¶

Setting	Default	Purpose
`ADMIN_ENABLED`	`True`	Mount the admin at all
`ADMIN_URL`	`"/admin"`	Where to mount it
`ADMIN_SITE_HEADER`	`"mlango administration"`	Header text
`ADMIN_SITE_TITLE`	`"mlango"`	Browser title
`ADMIN_PAGE_SIZE`	`25`	Default rows per page
`ADMIN_USERNAME`	`"admin"`	Basic auth username
`ADMIN_PASSWORD`	`""`	Set it to require Basic auth

Serving¶

Setting	Default	Purpose
`ROOT_ROUTECONF`	`None`	Module holding `urlpatterns`
`SERVE_MIDDLEWARE`	request logging	Middleware, outermost first
`SERVE_HOST`	`"127.0.0.1"`	Bind address
`SERVE_PORT`	`8000`	Port
`SERVE_API_KEYS`	`[]`	Keys accepted by `ApiKeyMiddleware`
`SERVE_BLOCKED_TERMS`	`[]`	Terms rejected by `GuardrailMiddleware`

Logging¶

Setting	Default
`LOG_LEVEL`	`"INFO"`
`LOG_FORMAT`	`"%(asctime)s %(levelname)-7s %(name)s: %(message)s"`

Per-environment settings¶

The usual pattern is a base module plus overrides:

myproject/settings/base.py

INSTALLED_APPS = ["reviews", "support"]
METASTORE = {"URL": "sqlite:///mlango.db"}

myproject/settings/production.py

import os

from myproject.settings.base import *  # noqa: F403

DEBUG = False
SECRET_KEY = os.environ["MLANGO_SECRET_KEY"]
ADMIN_PASSWORD = os.environ["MLANGO_ADMIN_PASSWORD"]
METASTORE = {"URL": os.environ["DATABASE_URL"]}
STORAGE = {"BACKEND": "mlango.storage.s3.S3Storage", "ROOT": "s3://bucket/mlango"}
DEFAULT_PROVIDER = "anthropic"
SERVE_API_KEYS = os.environ["MLANGO_API_KEYS"].split(",")

python manage.py migrate --settings myproject.settings.production

Nested dict settings merge with their defaults, so overriding METASTORE["URL"] does not mean restating ECHO and POOL_PRE_PING.

Configuring without a module¶

Used by tests, scripts and anything without a manage.py. For a notebook, mlango.notebook() wraps this with defaults chosen for one.

from mlango.conf import settings

settings.configure(
    BASE_DIR="/tmp/scratch",
    METASTORE={"URL": "sqlite:///scratch.db"},
    DEFAULT_PROVIDER="echo",
    INSTALLED_APPS=[],
)

import mlango
mlango.setup()

Checking your configuration¶

python manage.py check

Reports the resolved settings module, installed apps, what is declared, the metastore URL and whether its tables exist, which trainers and providers are importable, whether every dotted path in DEFAULT_CALLBACKS, SERVE_MIDDLEWARE and STORAGE resolves, pending migrations, and whether the admin is authenticated.

Resolving those paths up front matters: a typo in DEFAULT_CALLBACKS would otherwise surface as the same import error repeated once per sweep trial.