diwire

Type-driven dependency injection for Python. Zero dependencies. Zero boilerplate.

diwire is a dependency injection container for Python 3.10+ that builds your object graph from type hints. It supports scopes + deterministic cleanup, async resolution, open generics, fast steady-state resolution via compiled resolvers, and free-threaded Python (no-GIL) — all with zero runtime dependencies.

Frameworks & integrations

• Web: FastAPI, Litestar, aiohttp, Starlette, Flask, Django
• Tasks: Celery
• Testing: pytest plugin (Injected parameters)
• Config: pydantic-settings
• Overview: Integrations

Installation

uv add diwire

Why diwire

• Zero runtime dependencies: easy to adopt anywhere. (Why diwire)
• Scopes + deterministic cleanup: generator/async-generator providers clean up on scope exit. (Scopes)
• Async resolution: aresolve() mirrors resolve() and async providers are first-class. (Async)
• Open generics: register once, resolve for many type parameters. (Open generics)
• Function injection: Injected[T] for ergonomic handlers. (Function injection)
• Framework/task support: request/job scope patterns for FastAPI, Litestar, aiohttp, Starlette, Flask, Django, and Celery. (Integrations)
• Named components + collect-all: Component("name") and All[T]. (Components)
• Concurrency + free-threaded builds: configurable locking via LockMode. (Concurrency)

Performance (benchmarked)

Benchmarks + methodology live in the docs: Performance.

In this benchmark suite on CPython 3.14.3 (Apple M3 Pro, strict mode):

• diwire is the top performer across this suite, reaching up to 6.89× vs rodi, 30.79× vs dishka, and 4.40× vs wireup.
• Resolve-only comparisons (scope-capable libraries): diwire reaches up to 3.64× (rodi), 4.14× (dishka), and 3.10× (wireup).
• Current benchmark totals: 11 full-suite scenarios and 5 resolve-only scenarios.

For quick local regression checks, run make benchmark (diwire-only). For full cross-library runs, use make benchmark-comparison (raw suite) or make benchmark-report / make benchmark-report-resolve (report artifacts).

Quick start (pure Python auto-wiring)

Define your classes. Resolve the top-level one. diwire figures out the rest.

from dataclasses import dataclass, field

from diwire import Container


@dataclass
class Database:
    host: str = field(default="localhost", init=False)


@dataclass
class UserRepository:
    db: Database


@dataclass
class UserService:
    repo: UserRepository

container = Container()
service = container.resolve(UserService)
print(service.repo.db.host)  # => localhost

Registration

Use explicit registrations when you need configuration objects, interfaces/protocols, cleanup, or multiple implementations.

Strict mode (opt-in):

from diwire import Container, DependencyRegistrationPolicy, MissingPolicy

container = Container(
    missing_policy=MissingPolicy.ERROR,
    dependency_registration_policy=DependencyRegistrationPolicy.IGNORE,
)

Container() enables recursive auto-wiring by default. Use strict mode when you need full control over registration and want missing dependencies to fail fast.

from typing import Protocol

from diwire import Container, Lifetime


class Clock(Protocol):
    def now(self) -> str: ...


class SystemClock:
    def now(self) -> str:
        return "now"


container = Container()
container.add(
    SystemClock,
    provides=Clock,
    lifetime=Lifetime.SCOPED,
)

print(container.resolve(Clock).now())  # => now

Register factories directly:

from diwire import Container

container = Container()


def build_answer() -> int:
    return 42

container.add_factory(build_answer)

print(container.resolve(int))  # => 42

Scopes & cleanup

Use Lifetime.SCOPED for per-request/per-job caching. Use generator/async-generator providers for deterministic cleanup on scope exit.

from collections.abc import Generator

from diwire import Container, Lifetime, Scope


class Session:
    def __init__(self) -> None:
        self.closed = False

    def close(self) -> None:
        self.closed = True


def session_factory() -> Generator[Session, None, None]:
    session = Session()
    try:
        yield session
    finally:
        session.close()


container = Container()
container.add_generator(
    session_factory,
    provides=Session,
    scope=Scope.REQUEST,
    lifetime=Lifetime.SCOPED,
)

with container.enter_scope() as request_scope:
    session = request_scope.resolve(Session)
    print(session.closed)  # => False

print(session.closed)  # => True

Function injection

Mark injected parameters as Injected[T] and wrap callables with @resolver_context.inject.

from diwire import Container, Injected, resolver_context


class Service:
    def run(self) -> str:
        return "ok"


container = Container()
container.add(Service)


@resolver_context.inject
def handler(service: Injected[Service]) -> str:
    return service.run()


print(handler())  # => ok

Static typing note: without a checker plugin, injected wrappers keep return types but accept permissive arguments. For precise mypy signatures (optional injected params, strict non-injected params, optional diwire_resolver kwarg), enable:

[tool.mypy]
plugins = ["diwire.integrations.mypy_plugin"]

Named components

Use Annotated[T, Component("name")] when you need multiple registrations for the same base type. For registration ergonomics, you can also pass component="name" to add_* methods.

from typing import Annotated, TypeAlias

from diwire import All, Component, Container


class Cache:
    def __init__(self, label: str) -> None:
        self.label = label


PrimaryCache: TypeAlias = Annotated[Cache, Component("primary")]
FallbackCache: TypeAlias = Annotated[Cache, Component("fallback")]


container = Container()
container.add_instance(Cache(label="redis"), provides=Cache, component="primary")
container.add_instance(Cache(label="memory"), provides=Cache, component="fallback")

print(container.resolve(PrimaryCache).label)  # => redis
print(container.resolve(FallbackCache).label)  # => memory
print([cache.label for cache in container.resolve(All[Cache])])  # => ['redis', 'memory']

Resolution/injection keys are still Annotated[..., Component(...)] at runtime.

resolver_context (optional)

If you can't (or don't want to) pass a resolver everywhere, use resolver_context. It is a contextvars-based helper used by @resolver_context.inject and (by default) by Container resolution methods. Inside with container.enter_scope(...):, injected callables resolve from the bound scope resolver; otherwise they fall back to the container registered as the resolver_context fallback (Container(..., use_resolver_context=True) is the default).

from contextvars import ContextVar

from diwire import Container, Injected, Scope, resolver_context

current_user_id_var: ContextVar[int] = ContextVar("current_user_id", default=0)


def read_current_user_id() -> int:
    return current_user_id_var.get()


container = Container()
container.add_factory(read_current_user_id, provides=int, scope=Scope.REQUEST)


@resolver_context.inject(scope=Scope.REQUEST)
def handler(value: Injected[int]) -> int:
    return value


with container.enter_scope(Scope.REQUEST) as request_scope:
    token = current_user_id_var.set(7)
    try:
        print(handler(diwire_resolver=request_scope))  # => 7
    finally:
        current_user_id_var.reset(token)

Stability

diwire targets a stable, small public API.

• Backward-incompatible changes only happen in major releases.
• Deprecations are announced first and kept for at least one minor release (when practical).

Docs

• Tutorial (runnable examples)
• Web frameworks
• FastAPI quick start
• Examples (repo)
• Core concepts
• API reference

License

MIT. See LICENSE.