Korvex

A fast router for Python, written in Rust, with an ASGI app on top and a built-in server to run it.

Korvex is young (alpha). It's fast (see Performance), but it lacks some of what full frameworks offer: validation, OpenAPI. The roadmap lists what's coming.

pip install korvex

The router

Router maps a path and an HTTP method to a handler name. It never calls your code: what you do with the name is up to you.

from korvex import Router

router = Router()
router.get("/users/{id:int}", "get_user")
router.post("/users", "create_user", middleware=["auth"])

router.match("/users/42")         # ("get_user", {"id": 42}, ())
router.match("/users", "POST")    # ("create_user", {}, ("auth",))
router.match("/nope")             # None
router.match("/users", "DELETE")  # raises MethodNotAllowed, .allowed == ["POST"]

Path syntax

{id:int} gives an int and {id:uuid} a uuid.UUID; every other captured value is a string. When several routes could match, the most specific one wins: literal, then constrained parameters (in the order their routes were registered), then {id}, then *path.

Define your own constraints as regexes, before the routes that use them:

router.constraint("slug", r"[a-z0-9]+(?:-[a-z0-9]+)*")   # app.constraint() for an App
router.get("/posts/{post:slug}", "get_post")             # /posts/hello-world, not /posts/Hello

The whole segment must match. The regex is compiled and matched in Rust (the regex crate: no look-around or backreferences, so matching takes linear time), and the router still only gets a path and gives back a name. A constraint can't be redefined differently once routes may use it. In an App, the value is a string; an annotation converts it further (day: date on {day:isodate}), and OpenAPI shows the regex.

API

Invalid paths or methods raise ValueError.

match keeps the last 1,024 distinct lookups in a cache, and adding or removing a route clears it. Its results are shared between calls, so they are immutable: params is a read-only mapping and middleware a tuple.

Route groups

api = router.group("/api", middleware=["auth"])
api.get("/users", "list_users")             # GET /api/users, middleware ("auth",)

admin = api.group("/admin", middleware=["admin"])
admin.get("/stats", "stats")                # GET /api/admin/stats, middleware ("auth", "admin")

The ASGI app

App runs the matched handler for you. Serve it with any ASGI server (uvicorn, hypercorn, granian...).

from korvex import App, Depends, Header, HTTPException, Response

app = App()

def signed_in(authorization: str = Header("")) -> None:
    if not authorization:
        raise HTTPException(401, "sign in first")  # stops here

@app.get("/users/{id:int}", dependencies=[Depends(signed_in)])
async def get_user(request, id):
    return Response.json({"id": id})

pip install "korvex[serve]"   # installs uvicorn
uvicorn mymodule:app

• Handlers receive the Request plus path params as keyword arguments, and return a Response, sync or async. Checks shared by several routes (signing in, permissions) are dependencies: see Dependencies.
• Request has method, path, headers, query_params, path_params, cookies, client ((host, port)), scheme, and await request.body(), await request.json(), await request.form(). A form maps each field to its values; files from a multipart form are UploadedFiles (filename, content_type, data). A body that can't be parsed is answered with a 400.
• Response(body, status=, headers=), Response.text(...), Response.json(...), Response.redirect(...), and response.set_cookie(...) / response.delete_cookie(...).
• Response.stream(chunks) sends a body piece by piece, from a sync or async generator. Response.file(path, filename=None) sends a file in chunks (with filename, as a download). app.static("/static", "public") serves a directory; paths that leave it get a 404.
• HEAD requests are answered by the GET route, without the body.
• Unknown paths get a 404; known paths with the wrong method get a 405 with an Allow header.
• @app.before_request runs before every request, ahead of routing (so for unknown paths too), and returns a Response to answer right away or None to carry on. It runs for WebSocket handshakes too, with the WebSocket as the request: answering refuses the connection. @app.after_request gets every response, errors included, and returns the one to send (or None to keep it).
• raise HTTPException(404, "user not found") in a handler, a dependency or a before hook answers with that error; raise Abort(response) with that exact response (a redirect to the sign-in page, say), past the exception handlers. @app.exception_handler(404) (a status) or @app.exception_handler(ValueError) (an exception class) customizes the response. Any other exception gives a 500 and is logged to the korvex logger.
• app.group(prefix, dependencies=[...]) works like router.group(), with decorators: api = app.group("/api"), then @api.get("/users"). Groups nest; a group's dependencies run before its routes' own.
• app.host("api.example.com") gives a group whose routes only answer for that host (its Host header, port aside); app.host("{tenant}.example.com") passes tenant to the dependencies and handlers that take it, like a path parameter. The first host pattern that fits owns the request (those without parameters first): its routes answer, then the routes declared without a host, which answer on every host. /docs documents the routes of the host it's served from, and app.openapi(host=...) those of any host. In the built-in server, a route's max_body applies to its path on every host.
• Routes registered on app.router directly are linked to a function with @app.handler("name"). The App doesn't run the router's middleware names: a route that has some answers a 500, rather than run without what may be its guard.
• @app.on_startup and @app.on_shutdown run a function (sync or async) when the server starts and stops, for example to open and close a database pool. Shutdown hooks run in reverse order. If a startup hook raises, the server doesn't start. Both @app.on_startup and @app.on_startup() work.

Validation

Type annotations say where a handler's parameters come from. Models can be msgspec structs or pydantic models; Korvex needs neither.

import msgspec

class UserIn(msgspec.Struct):
    name: str
    age: int

class UserOut(msgspec.Struct):
    id: int
    name: str

@app.post("/users")
async def create(request, body: UserIn, notify: bool = False) -> UserOut:
    return UserOut(id=1, name=body.name)

• A model (or list[Model]) parameter is the JSON body, validated.
• A path parameter's name gets its value, converted to the annotated type.
• Other str, int, float, bool, UUID, date or datetime parameters (or Optional or list of them) come from the query string, dates in ISO 8601; a default makes them optional. Any other annotation is refused when the route is registered.
• A model return annotation turns the returned object into JSON; a returned Response is sent as is.
• Invalid input gets a 422 with {"detail": [{"loc": ["body", "age"], "msg": ..., "type": ...}]}. pydantic lists every problem; msgspec stops at the first. @app.exception_handler(422) changes the response.
• Annotations are read once, when the route is registered: a handler without them costs nothing extra, and one that can't be understood fails at import.

Headers, cookies and form fields are read the same way, with Header(), Cookie() and Form() as the default (or in Annotated[...]):

from korvex import Cookie, Form, Header, Query, UploadedFile

@app.post("/photos")
async def upload(
    request,
    title: str = Form(),                         # a form field, urlencoded or multipart
    photo: UploadedFile = Form(),                # a file from a multipart form
    x_client: str = Header("web"),               # the X-Client header
    theme: str = Cookie("light"),                # the theme cookie
    page_size: int = Query(20, alias="page-size"),
):
    ...

• x_client reads X-Client: underscores become hyphens, case doesn't matter. alias= gives the exact name.
• Their errors say where: ["header", "x-client"], ["form", "title"].
• An empty form field counts as missing, except for str.
• A model can come from a form too, field by field: body: Signup = Form(). A list field takes every value, a checkbox's on is true, and text converts to numbers and dates as the model says. Files stay separate Form() parameters.

Dependencies

Depends(func) gives a parameter the value func returns. func takes parameters the way a handler does (query, headers, cookies, body, path, other dependencies), plus the request if it asks for it:

from typing import Annotated, Optional
from korvex import Cookie, Depends, HTTPException

async def current_user(session: Optional[str] = Cookie(None)) -> User:
    user = await users.by_session(session)
    if user is None:
        raise HTTPException(401, "sign in first")
    return user

CurrentUser = Annotated[User, Depends(current_user)]

async def database():
    connection = await pool.acquire()
    try:
        yield connection                     # the handler runs here
    finally:
        await pool.release(connection)

@app.get("/tasks")
async def tasks(request, user: CurrentUser, db=Depends(database)) -> list[Task]:
    return await db.tasks_of(user.id)

api = app.group("/api", dependencies=[Depends(current_user)])  # checks every route in it

• A dependency can be sync or async, and stop the request by raising HTTPException. Within one request it runs once, however many parameters ask for it (Depends(func, cache=False) runs it each time).
• One that yields is torn down when the handler is done (for a streamed response, once the stream is sent). An error from the handler is raised at the yield, so try/except/finally work there; the error still gets its usual response.
• dependencies=[...] on a route, a group or a WebSocket runs them for what they check rather than what they return. A group's run before its routes' own.
• In tests, app.dependency_overrides[current_user] = lambda: test_user replaces one everywhere.
• Validation errors from dependencies and from the handler come in one 422.

app.openapi(title=..., version=...) describes the API as an OpenAPI 3.1 document: paths, typed parameters (path, query, headers, cookies, those of dependencies included), JSON and form bodies, response models, and each handler's docstring (its first line as the summary). app.docs(title=..., version=...) serves it at /openapi.json, with Swagger UI at /docs to try the API from a browser (loaded from a CDN; docs_path=None leaves it out).

WebSockets

@app.websocket("/ws/{room}")
async def chat(ws, room, user: CurrentUser):  # no session: refused (403)
    await ws.accept()
    async for message in ws:  # text or bytes, until the client leaves
        await ws.send_text(f"{user.name} in {room}: {message}")

• A WebSocket has the handshake's headers, cookies, query_params, path_params, client and scheme; accept(subprotocol=...), receive() (raises WebSocketDisconnect), receive_text/bytes/json, send_text/bytes/json and close(code).
• Parameters and dependencies work as for HTTP handlers, without a body. A dependency raising HTTPException refuses the connection before it opens (403); a handler can also close() before accept() to refuse.
• They work the same under ASGI servers and in the built-in server. Groups support them (api.websocket(...)), group dependencies included.
• Before hooks run on the handshake, so request_id() and rate_limit() apply, and a hook can refuse a connection whose Origin isn't yours: browsers don't check it for WebSockets. After hooks don't run: there's no response to change.

Ready-made hooks

korvex.middleware registers hooks for common jobs:

from korvex.middleware import cors

cors(app, allow_origins=["https://example.com"], allow_credentials=True)

• cors(app, allow_origins=..., allow_methods=..., allow_headers=..., allow_credentials=False, expose_headers=(), max_age=600) lets browsers call the app from other origins. Preflight requests are answered before routing; every response to an allowed origin, errors included, gets the CORS headers.
• request_id(app, header="X-Request-ID", trust_incoming=True) gives every request an ID (the incoming one if it's safe, a new one otherwise) and sends it back. current_request_id() returns it anywhere in your code, RequestIdFilter adds it to log records (%(request_id)s), and the built-in server's access logs show it.
• rate_limit(app, requests=100, per=60, key=...) allows each key (the client's address by default, or an API key, a user ID...) that many requests per period; beyond that, a 429 with Retry-After. Responses carry RateLimit-Limit, RateLimit-Remaining and RateLimit-Reset. Counters live in each worker's memory, so with 3 workers a client can get up to 3 times the limit; store=RedisStore("redis://...") shares them between workers and machines (pip install korvex[redis]). If Redis can't be reached, requests go through (fail_open=False refuses them), and the error is logged.
• compress(app) compresses text-like responses over 500 bytes with zstd (Python 3.14), brotli (pip install korvex[brotli]) or gzip, whichever the client prefers among those it accepts, in that order. Streamed responses are compressed piece by piece, each flushed, so a live stream isn't held back. Call it before the other hooks. Mind BREACH: don't compress pages that hold a secret next to something the client sent.
• security_headers(app) adds X-Frame-Options: DENY, X-Content-Type-Options: nosniff, Referrer-Policy, Cross-Origin-Opener-Policy, and Strict-Transport-Security on HTTPS requests; content_security_policy= adds a CSP (off by default: the right one depends on your pages, and /docs loads Swagger UI from a CDN). A response that sets one of these keeps its own.

Behind a reverse proxy, let it compress and add headers instead: it does both well, and the app does less.

More in examples/, up to a complete app laid out like a real project. To put it in production (behind nginx or Caddy, in Docker, on AWS Lambda or Vercel), see the deployment guide.

The built-in server

Korvex can run an App without an ASGI server, on Linux and macOS. On the same CPU cores, it serves 20 to 36% more requests than uvicorn with about 40% less memory (see Performance).

korvex run main:app --workers 3

or, from Python, app.run(workers=3). The same options exist in both:

With threads, a module-level global set in a startup hook would hold one loop's connection pool for all of them, and asyncio objects only work on the loop that made them. app.local keeps attributes per event loop:

@app.on_startup
async def connect():
    app.local.pool = await asyncpg.create_pool(DATABASE_URL)   # this loop's pool

@app.get("/users/{id:int}")
async def get_user(request, id):
    row = await app.local.pool.fetchrow("SELECT ... WHERE id = $1", id)

With one loop, it's simply an attribute, under any server.

How it works: a Rust thread per worker reads each request in full, without touching Python, then hands it to the asyncio loop, which runs your handlers. A slow or huge upload never holds a Python worker.

Things to know:

• Run it behind a reverse proxy (nginx, Caddy, Traefik) in production: it speaks HTTP/1.1 and WebSocket; HTTPS, HTTP/2 and HTTP/3 are the proxy's job.
• A body over the limit (413), too slow (408) or broken (400) is decided by the server, but answered by the app, without the body: before hooks, @app.exception_handler(413), after hooks and the access log all apply. The connection closes after it.
• A slow client can't make a stream or a WebSocket pile up in memory: once 1 MiB waits for it, the handler's next send waits until it catches up.
• A proxy passes on the headers it doesn't set as the client wrote them. So Forwarded (RFC 7239) is only read when there's no X-Forwarded-For or -Proto, which nginx, Caddy and Traefik set; the forwarded host only with --forwarded-host; and for -Proto and -Host, the rightmost value counts, the one the nearest proxy set.
• The App is still a plain ASGI app: uvicorn, Granian or a serverless platform keep working, and Windows needs one of them.

Performance

Servers

The same app, served by each server on the same 4 CPU cores, each with its best number of workers:

• Granian is a Rust ASGI server too: it does better than uvicorn here, but the built-in server still leads, because it skips the ASGI layer.
• Robyn runs its own version of the app with its own Rust server.
• No request failed.

Reproduce with uv run python benches/servers.py (needs cargo install oha, takes about 4 minutes).

Under load

The same small app (1,000 routes, one JSON endpoint) in each framework, served the same way (uvicorn, one worker, one CPU core) and flooded with requests to random URLs by oha:

• Every framework uses a full CPU core under load, so requests per second is also how much work each one gets out of that core.
• Falcon, and Starlette in its best case, come within about 10% of Korvex.
• Starlette, FastAPI and Django try routes one by one: the ranges go from the tested route declared after the 1,000 others to declared first. BlackSheep tries them one by one too, but sorts them itself, deepest first, so here the tested route always comes last.
• No request failed. Throughput holds steady as clients are added; latency grows because requests queue, not because each one gets slower. After the flood, memory stays flat and speed is back to normal.

Robyn isn't in this table, since it can't run on uvicorn: it's compared with Korvex's built-in server above.

Reproduce with uv run python benches/load_test.py (needs cargo install oha, takes about 11 minutes).

Router alone

Route matching only, without any server:

Median time per lookup, 10,000 routes:

• Korvex is the fastest in both cases, and its time stays flat as routes are added. Django and Starlette try routes one by one, so they slow down linearly.
• Korvex, Litestar and Sanic cache the last 1,024 lookups. The cache helps most with static URLs; URLs with parameters (/users/1, /users/2, ...) are often different, so they hit it less.
• With few routes, the server and the handler cost more than the router. With many routes, a router that tries them one by one becomes the main cost (see Under load).

Reproduce with uv run pytest benches/bench_router.py --benchmark-only --benchmark-json=benches/results.json, then uv run python benches/plot.py to redraw both charts. Results depend on the machine: measure your own workload before relying on any number.

Validation

One POST with a small JSON body, answered with JSON, run through the app in-process (no server). Median time per request:

• Validating is faster than not validating here: msgspec and pydantic decode and encode JSON in compiled code, request.json() and Response.json() use Python's json module.
• Handlers without annotations take the same time as in 0.6, validated ones the same as in 0.7 (measured side by side).

Reproduce with uv run pytest benches/bench_validation.py --benchmark-only.

Dependencies

One GET whose handler gets the user from a header and opens then closes a session, run through the app in-process. Median time per request, from one run on the same machine:

• FastAPI runs sync dependencies in a thread pool, which keeps a blocking one from stopping the event loop but costs a lot. Korvex calls them on the loop, like sync handlers: keep blocking work out of them, or make them async.
• Dependencies are worked out when the route is registered; a handler without them pays nothing.

Reproduce with uv run pytest benches/bench_dependencies.py --benchmark-only.

Threads on free-threaded Python

The built-in server on the same 4 cores as above, the same app and load (200 connections), with more threads or more workers. One run:

• On 3.14t, 3 threads serve 57% more than one, within 4% of 4 workers, with a third of the memory. Past 3, they don't gain: each loop has its Rust thread too, so 4 of them already keep 8 threads busy on 4 cores.
• With the GIL, threads don't help: use workers.
• Workers still serve the most; threads are for when memory counts (one copy of the app, its caches and connection pools per process).
• compress() leaves brotli out on 3.14t: the brotli package turns the GIL back on for the whole process. CI checks that nothing else does.

Reproduce with KORVEX_FT_PYTHON=/path/to/3.14t/python uv run python benches/threads.py (see the file for the setup).

With a database

Each request reads one row from PostgreSQL through an asyncpg pool, then answers with JSON: the least a real API does. PostgreSQL gets 2 cores of its own, the server 4, each server 3 workers. Two runs:

• The database sets the pace: Korvex's two servers serve about the same, uvicorn a little more in the second run, and the built-in server's lead of the tests above is gone. It still uses the least memory.
• FastAPI serves 35 to 40% less: its own cost per request comes on top of the database's.

Reproduce with uv run python benches/database.py (needs podman or docker for PostgreSQL, and oha).

Development

uv run maturin develop --release   # build the Rust extension
uv run pytest                      # Python tests
cargo test                         # Rust tests

See CONTRIBUTING.md, and ROADMAP.md for what's planned.

License

MIT