Weft

Weft is an ASGI and WSGI server for FastAPI, Flask and the rest. It is written in Rust on Tokio and runs inside the Python thread that runs your application: there is no second runtime thread, no queue between Rust and Python, and no handoff per request. The HTTP server and asyncio share one thread, one event loop, and one wakeup.

weft app:app
weft app:app --workers 4
weft app:app --workers 4 --worker-mode thread     # free-threaded Python (3.14t)

import weft

weft.run(app, host='0.0.0.0', port=8000)

HTTP/1.1, HTTP/2, HTTP/3, WebSocket and WebTransport. TLS, static files, compression, a response cache, and Prometheus. ACME is left to Caddy.

Further reading: ARCHITECTURE.md — how it is built. BENCHMARKS.md — Weft, Peregrine and Granian on the-benchmarker contract. RELEASE.md — what changed. bench/README.md — the closed-loop FastAPI matrix.

How it works

• Tokio is driven from asyncio's selector. WeftEventLoop is a SelectorEventLoop whose selector is a Tokio LocalRuntime. Every select() runs the runtime until a Python-side file descriptor is ready, a timer is due, or the application has work to do. asyncio's own sockets, subprocesses and call_soon_threadsafe keep working; they are registered with the same Tokio driver. The GIL is released only while Tokio parks.
• The ASGI interface is C, not Python. send and receive are native vectorcall objects. The scope is a copy of a prebuilt dict with interned keys. Header names are cached. await send(...) writes to the socket before it returns, and hands back an already-completed awaitable unless the client is slow to read (then a real future provides backpressure). A body that has already arrived is returned by await receive() without suspending.
• HTTP/1.1 in Rust: httparse for heads, keep-alive and pipelining, content-length and chunked bodies both ways, Expect: 100-continue, HEAD/204/304 semantics, request smuggling checks (TE+CL, invalid lengths), header timeouts, a head size limit, and graceful shutdown that finishes in-flight requests.
• WSGI too. A two-argument application (Flask, Django's WSGI handler, anything PEP 3333) is detected and served the same way; --protocol wsgi or asgi overrides the guess. The application is called inline once its body has been read; start_response is a native object that doubles as write(), and the response iterable is sent block by block, parking the connection on the socket between blocks so a slow client holds up only its own generator. --wsgi-threads N runs the application on N threads per worker instead, for blocking handlers; the worker thread still does all the socket work.
• Operations, in the server: an access log (--access-log, text or --access-log-format json) buffered per worker and written in batches; --health-check-path answered without Python; --drain-delay for load balancers; --max-connections; --forwarded-allow-ips for X-Forwarded-* and Forwarded; --request-id, --trace-context and --request-start-header for tracing and queue-time APMs; --rate-limit N/s counted across every worker; --unix / --uds; --reload and SIGHUP replace process workers one at a time; --venv / --no-auto-venv; --no-lifespan; --static-dir PREFIX=DIR; --compress / --compress-static; --cache-size shared by every worker; --tls-cert / --tls-key (SNI when repeated), --hsts, --redirect-http; HTTP/2 by default (--no-http2, --http2-only); HTTP/3 over QUIC (--http3, --quic-port) and WebTransport (ASGI webtransport.*: streams and datagrams); Prometheus on --metrics-port / --metrics-host.
• Workers: one process by default; --workers N starts N processes that share the listening socket, or N threads with --worker-mode thread, which on free-threaded Python run in parallel in one process. Each worker is a complete, independent loop.

Performance

the-benchmarker contract, 256 connections, 4 workers, zrk open-loop ramp, CPython 3.14.7. Median of three 15 s runs, every request 2xx. How that was measured.

Weft is 1.5–1.9× Granian on this machine. Raw ASGI and FastAPI match Peregrine; WSGI does not — Peregrine's WSGI path has no asyncio.

Closed loop on Windows 11, one worker, 64 connections (python bench/matrix.py --set fastapi), requests per second:

Each server uses one core. With four worker processes (--set scaling), GET / reaches 35,200 requests per second against uvicorn's 24,300, and the path parameter endpoint 26,800 against 20,100.

Install

pip install weft-server

Distribution name: weft-server (the PyPI name weft belongs to an unrelated project). Import package weft, CLI weft. CPython 3.10+; wheels for Linux, macOS and Windows. Tested on 3.14 and free-threaded 3.14t.

From source:

uv venv --python 3.14 .venv
uv pip install maturin
maturin develop --release

Scope

Supported: ASGI 3.0 HTTP (spec 2.4), WebSocket and lifespan; WSGI (PEP 3333); HTTP/1.0, 1.1, HTTP/2 (h2c prior knowledge, ALPN over TLS) and HTTP/3 over QUIC; WebTransport (ASGI webtransport.*); TCP and unix sockets; TLS (rustls). ACME is left to the reverse proxy (Caddy).

weft.contrib has the framework helpers a WebTransport or HTTP/3 app usually writes once: WebTransportRouter (Starlette {name} / Django <int:pk> paths), WebTransportEndpoint, AltSvcMiddleware, weft.webtransport.WebTransportSession, and IETF resumable uploads (ResumableUploads). Starlette, FastAPI and Django serve HTTP/3 unchanged; the routers exist because those frameworks assert on scope["type"] before they would see a webtransport session.

Tests

pytest tests

The suite runs raw ASGI and WSGI apps and a FastAPI app over real sockets, in process, thread and inline worker modes, on both the standard and the free-threaded interpreter.