gi0baro/tonio

A multi-threaded async runtime for Python

Python

168

240 commits

updated Oct 2, 2026

See the code

README

TonIO

TonIO is a multi-threaded async runtime for free-threaded Python, built in Rust on top of the mio crate, and inspired by tinyio, trio and tokio.

Warning: TonIO is currently a work in progress. The APIs are subject to breaking changes.

Note: TonIO is available on free-threaded Python only. Windows is supported on a best-effort basis.

TonIO supports both using yield and the more canonical async/await notations, with the latter being available as part of the tonio.colored module. Following code snippets show both the usages.

Warning: despite the fact TonIO supports async and await notations, it's not compatible with any asyncio object like futures and tasks. The TonIO-Monkey project provides patches for some popular asyncio packages.

In a nutshell

yield syntax

import tonio

def wait_and_add(x):
    yield tonio.sleep(1)
    return x + 1

@tonio.main
def main():
    parallel = tonio.spawn(
        wait_and_add(3), 
        wait_and_add(4)
    )
    six = yield wait_and_add(5)
    four, five = yield parallel
    return four, five, six

assert main() == (4, 5, 6)

await syntax

import tonio.colored as tonio

async def wait_and_add(x):
    await tonio.sleep(1)
    return x + 1

@tonio.main
async def main():
    parallel = tonio.spawn(
        wait_and_add(3), 
        wait_and_add(4)
    )
    six = await wait_and_add(5)
    four, five = await parallel
    return four, five, six

assert main() == (4, 5, 6)

Usage

Entrypoint

Every TonIO program consists of an entrypoint, which should be passed to the run method:

yield syntax

import tonio

def main():
    yield
    print("Hello world")

tonio.run(main())

await syntax

import tonio.colored as tonio

async def main():
    await tonio.yield_now()
    print("Hello world")

tonio.run(main())

TonIO also provides a main decorator, thus we can rewrite the previous example as:

yield syntax

import tonio

@tonio.main
def main():
    yield
    print("Hello world")

main()

await syntax

import tonio.colored as tonio

@tonio.main
async def main():
    await tonio.yield_now()
    print("Hello world")

main()

Note: as you can see the colored module provides the additional yield_now coroutine, a quick way to define a suspension point, given you cannot just yield as in the non-colored notation.

Note: both run and main can only be called once per program. To run the runtime multiple times in the same program, follow the section below.

Manually managing the runtime

TonIO also provides the runtime function, to manually manage the runtime lifecycle:

import tonio

def _run1():
    ...

async def _run2():
    ...

def main():
    runtime = tonio.runtime()
    runtime.run_until_complete(_run1())
    runtime.run_until_complete(_run2())

Runtime options

The run, main and runtime methods accept options, specifically:

option name	description	default
`context`	enable `contextvars` usage in coroutines	`False`
`signals`	list of signals to listen to
`threads`	Number of runtime threads	# of CPU cores
`blocking_threadpool_size`	Maximum number of blocking threads	128
`blocking_threadpool_idle_ttl`	Idle timeout for blocking threads (in seconds)	30

Events

The core object in TonIO is Event. It's basically a wrapper around an atomic boolean flag, initialised with False. Event provides the following methods:

is_set(): return the value of the flag
set(): set the flag to True
clear(): set the flag to False
wait(timeout=None): returns a coroutine you can yield on that unblocks when the flag is set to True or the timeout expires. Timeout is in seconds.

yield syntax

import tonio

@tonio.main
def main():
    event = tonio.Event()

    def setter():
        yield tonio.sleep(1)
        event.set()

    tonio.spawn(setter())
    yield event.wait()

await syntax

import tonio.colored as tonio

@tonio.main
async def main():
    event = tonio.Event()

    async def setter():
        await tonio.sleep(1)
        event.set()

    tonio.spawn(setter())
    await event.wait()

Waiters

Event.wait returns a Waiter, the object the runtime actually suspends on. Waiters can be combined: w1 & w2 unblocks when both are set, w1 | w2 when either is. When the operands carry timeouts, & keeps the longest and | the shortest.

Waiters can also be built directly from events with Waiter(ev1, ev2) (all) and Waiter.any(ev1, ev2). Both accept a timeout that applies to the waiter as a whole, and expressed in microseconds.

yield syntax

import tonio

@tonio.main
def main():
    ready, stop = tonio.Event(), tonio.Event()
    yield ready.wait() | stop.wait()
    if stop.is_set():
        return

await syntax

import tonio.colored as tonio

@tonio.main
async def main():
    ready, stop = tonio.Event(), tonio.Event()
    await (ready.wait() | stop.wait())
    if stop.is_set():
        return

Result

Result is a thread-safe slot to hand values across coroutines, typically paired with an Event. store(value) writes it and fetch() reads it back. Built with a size, it holds that many slots, store(value, index) fills one and fetch() returns them as a list.

Spawning tasks

TonIO provides the spawn method to schedule new coroutines onto the runtime:

yield syntax

import tonio

def doubv(v):
    yield
    return v * 2

@tonio.main
def main():
    parallel = tonio.spawn(doubv(2), doubv(3))
    v3 = yield doubv(4)
    v1, v2 = yield parallel
    print([v1, v2, v3])

await syntax

import tonio.colored as tonio

async def doubv(v):
    await tonio.yield_now()
    return v * 2

@tonio.main
async def main():
    parallel = tonio.spawn(doubv(2), doubv(3))
    v3 = await doubv(4)
    v1, v2 = await parallel
    print([v1, v2, v3])

Coroutines passed to spawn get scheduled onto the runtime immediately. Using yield or await on the return value of spawn just waits for the coroutines to complete and retrieve the results.

When results are not needed, spawn.without_results waits for completion without collecting them. spawn.without_tracking schedules the coroutines and returns nothing at all, for fire-and-forget work.

Blocking tasks

TonIO provides the spawn_blocking method to schedule blocking operations onto the runtime:

yield syntax

import tonio

def read_file(path):
    with open(path, "r") as f:
        return f.read()

@tonio.main
def main():
    file_data = yield tonio.spawn_blocking(
        read_file, 
        "sometext.txt"
    )

await syntax

import tonio.colored as tonio

def read_file(path):
    with open(path, "r") as f:
        return f.read()

@tonio.main
async def main():
    file_data = await tonio.spawn_blocking(
        read_file, 
        "sometext.txt"
    )

Running tasks from synchronous contexts

TonIO provides the block_on method to spawn coroutines from a synchronous context. It works the same way of spawn, except it accepts a single coroutine and it blocks the current thread until the coroutine is completed.

Warning: using block_on from within a coroutine might produce a runtime deadlock.

Map utilities

TonIO provides the map and map_blocking utilities to spawn the same operation with an iterable of parameters:

yield syntax

import tonio

accum = []

def task(no):
    yield tonio.sleep(0.5)
    accum.append(no * 2)

@tonio.main
def main():
    yield tonio.map(task, range(4))

await syntax

import tonio.colored as tonio

accum = []

async def task(no):
    await tonio.sleep(0.5)
    accum.append(no * 2)

@tonio.main
async def main():
    await tonio.map(task, range(4))

Completion-based iterators

TonIO provides the as_completed utility to iterate over task results based on completion order:

yield syntax

import tonio

def _sleep(v):
    yield tonio.sleep(v)
    return v

@tonio.main
def main():
    vals = []
    for task in tonio.as_completed(
        _sleep(0.5),
        _sleep(0.1),
        _sleep(0.3),
    ):
        vals.append(yield task)

await syntax

import tonio.colored as tonio

async def _sleep(v):
    await tonio.sleep(v)
    return v

@tonio.main
async def main():
    vals = []
    async for val in tonio.as_completed(
        _sleep(0.5),
        _sleep(0.1),
        _sleep(0.3),
    ):
        vals.append(val)

Scopes and cancellations

Note: cancellations in TonIO are optimistic and terminal; optimistic means that cancelling takes effect only at suspension points, terminal means that once a coroutine is cancelled every later suspension is cancelled as well.

Warning: some TonIO primitives are not cancel-safe; they are marked as such in their sections. An operation is considered cancel-safe when cancelling it leaves the objects it operates on untouched.

TonIO provides a scope context, that lets you cancel work spawned within it:

yield syntax

import tonio

def slow_push(target, sleep):
    yield tonio.sleep(sleep)
    target.append(True)

@tonio.main
def main():
    values = []
    with tonio.scope() as scope:
        scope.spawn(slow_push(values, 0.1))
        scope.spawn(slow_push(values, 2))
        yield tonio.sleep(0.2)
        scope.cancel()
    yield scope()
    assert len(values) == 1

await syntax

import tonio.colored as tonio

async def slow_push(target, sleep):
    await tonio.sleep(sleep)
    target.append(True)

@tonio.main
async def main():
    values = []
    async with tonio.scope() as scope:
        scope.spawn(slow_push(values, 0.1))
        scope.spawn(slow_push(values, 2))
        await tonio.sleep(0.2)
        scope.cancel()
    assert len(values) == 1

When you yield on the scope, it will wait for all the spawned coroutines to end. If the scope was canceled, then all the pending coroutines will be canceled. By default, an exception in the scope context won't cancel the scope itself. If you want those exceptions to cancel the scope, you can pass cancel_on_exc=True to scope.

Note: as you can see, the colored version of scope doesn't require to be awaited, as it will yield when exiting the context.

Select first completing task

TonIO also provides a select utility to cancel remaining work on the first completing task:

yield syntax

import tonio

def slow_push(target, sleep):
    yield tonio.sleep(sleep)
    target.append(True)

@tonio.main
def main():
    values = []
    yield tonio.select(
        slow_push(values, 0.1),
        slow_push(values, 2)
    )
    assert len(values) == 1

await syntax

import tonio.colored as tonio

async def slow_push(target, sleep):
    await tonio.sleep(sleep)
    target.append(True)

@tonio.main
async def main():
    values = []
    await tonio.select(
        slow_push(values, 0.1),
        slow_push(values, 2)
    )
    assert len(values) == 1

Note: select also accepts waiters, so an Event.wait() can race against coroutines.

tonio.time.time(): a function returning the runtime's clock (in seconds, microsecond resolution)
tonio.time.sleep(delay): a coroutine you can yield on to sleep (delay is in seconds)
tonio.time.timeout(coro, timeout): a coroutine you can yield on returning a tuple (output, success). If the coroutine succeeds in the given time then the pair (output, True) is returned. Otherwise this will return (None, False).

Note: time.sleep is also exported to the main tonio module.

Note: all of the above functions are also present in tonio.colored.time module.

Scheduling work

TonIO provides the time.interval function to create interval objects you can yield on a scheduled basis:

yield syntax

import tonio
from tonio import time

def some_task():
    ...

def scheduler():
    interval = time.interval(1)
    while True:
        yield interval.tick()
        tonio.spawn(some_task())

@tonio.main
def main():
    tonio.spawn(scheduler())
    # do some other work

await syntax

import tonio.colored as tonio
from tonio.colored import time

async def some_task():
    ...

async def scheduler():
    interval = time.interval(1)
    while True:
        await interval.tick()
        tonio.spawn(some_task())

@tonio.main
async def main():
    tonio.spawn(scheduler())
    # do some other work

The interval method first argument is the interval in seconds resolution, and the method also accepts an optional at argument, to delay the first execution at a specific time (from the runtime's clock perspective):

from tonio import time

# tick every 500ms, with the first tick happening in 5 seconds from now
interval = time.interval(0.5, time.time() + 5)

Warning: Interval.tick is not cancel-safe.

Synchronization primitives

Synchronization primitives are exposed in the tonio.sync module.

Lock

Implements a classic mutex, or a non-reentrant, single-owner lock for coroutines:

yield syntax

import tonio
from tonio import sync

@tonio.main
def main():
    # counter can't go above 1
    counter = 0

    def _count(lock):
        nonlocal counter
        with (yield lock()):
            counter += 1
            yield
            counter -= 1
    
    lock = sync.Lock()
    yield tonio.spawn(*[
        _count(lock)
        for _ in range(10)
    ])

await syntax

import tonio.colored as tonio
from tonio.colored import sync

@tonio.main
async def main():
    # counter can't go above 1
    counter = 0

    async def _count(lock):
        nonlocal counter
        async with lock:
            counter += 1
            await tonio.yield_now()
            counter -= 1
    
    lock = sync.Lock()
    await tonio.spawn(*[
        _count(lock)
        for _ in range(10)
    ])

The Lock object also implements an or_raise method, that will immediately fail when the lock cannot be acquired:

from tonio.exceptions import WouldBlock

try:
    with lock.or_raise():
        ...
except WouldBlock:
    ...

Semaphore

A semaphore for coroutines:

yield syntax

import tonio
from tonio import sync

@tonio.main
def main():
    # counter can't go above 2
    counter = 0

    def _count(semaphore):
        nonlocal counter
        with (yield semaphore()):
            counter += 1
            yield
            counter -= 1
    
    semaphore = sync.Semaphore(2)
    yield tonio.spawn(*[
        _count(semaphore)
        for _ in range(10)
    ])

await syntax

import tonio.colored as tonio
from tonio.colored import sync

@tonio.main
async def main():
    # counter can't go above 2
    counter = 0

    async def _count(semaphore):
        nonlocal counter
        async with semaphore:
            counter += 1
            await tonio.yield_now()
            counter -= 1
    
    semaphore = sync.Semaphore(2)
    await tonio.spawn(*[
        _count(semaphore)
        for _ in range(10)
    ])

As for locks, the Semaphore object also implements an or_raise method, that will immediately fail when the lock cannot be acquired:

from tonio.exceptions import WouldBlock

try:
    with semaphore.or_raise():
        ...
except WouldBlock:
    ...

The Semaphore object also implements a tokens method, that returns the number of available tokens.

Barrier

A barrier for coroutines:

yield syntax

import tonio
from tonio import sync

@tonio.main
def main():
    barrier = sync.Barrier(3)
    count = 0

    def _start_at_3():
        nonlocal count
        count += 1
        i = yield barrier.wait()
        assert count == 3
        return i

    yield tonio.spawn(*[
        _start_at_3()
        for _ in range(3)
    ])

await syntax

import tonio.colored as tonio
from tonio.colored import sync

@tonio.main
async def main():
    barrier = sync.Barrier(3)
    count = 0

    async def _start_at_3():
        nonlocal count
        count += 1
        i = await barrier.wait()
        assert count == 3
        return i

    await tonio.spawn(*[
        _start_at_3()
        for _ in range(3)
    ])

Warning: Barrier.wait is not cancel-safe.

The Barrier object also implements a value method, which returns the current value of the barrier.

Channels

Multi-producer multi-consumer channels for inter-coroutine communication.

The tonio.sync.channel module provides both a channel and an unbounded constructors.
The main difference between bounded and unbounded channels, as the names suggest, is that while the first will suspend sending messages once the specified length is reached, and it will resume accepting messages once the existing buffer is consumed, the latter will always accept new messages. That's also why, the sender part of a bounded channel is async, while in the unbounded is not.

Bounded channel

yield syntax

import tonio
from tonio import sync
from tonio.sync import channel

def producer(sender, barrier, offset):
    for i in range(20):
        message = offset + i
        yield sender.send(message)
    yield barrier.wait()

def consumer(receiver):
    while True:
        try:
            message = yield receiver.receive()
            print(message)
        except BrokenPipeError:
            break

@tonio.main
def main():
    def close(sender, barrier):
        yield barrier.wait()
        sender.close()

    sender, receiver = channel.channel(2)
    barrier = sync.Barrier(3)
    yield tonio.spawn(*[
        producer(sender, barrier, 100),
        producer(sender, barrier, 200),
        consumer(receiver),
        consumer(receiver),
        consumer(receiver),
        consumer(receiver),
        close(sender, barrier),
    ])

await syntax

import tonio.colored as tonio
from tonio.colored import sync
from tonio.colored.sync import channel

async def producer(sender, barrier, offset):
    for i in range(20):
        message = offset + i
        await sender.send(message)
    await barrier.wait()

async def consumer(receiver):
    while True:
        try:
            message = await receiver.receive()
            print(message)
        except BrokenPipeError:
            break

@tonio.main
async def main():
    async def close(sender, barrier):
        await barrier.wait()
        sender.close()

    sender, receiver = channel.channel(2)
    barrier = sync.Barrier(3)
    await tonio.spawn(*[
        producer(sender, barrier, 100),
        producer(sender, barrier, 200),
        consumer(receiver),
        consumer(receiver),
        consumer(receiver),
        consumer(receiver),
        close(sender, barrier),
    ])

Warning: receive and send operations are not cancel-safe in bounded channels.

Unbounded channel

yield syntax

import tonio
from tonio import sync
from tonio.sync import channel

def producer(sender, barrier, offset):
    for i in range(20):
        message = offset + i
        sender.send(message)
    yield barrier.wait()

def consumer(receiver):
    while True:
        try:
            message = yield receiver.receive()
            print(message)
        except BrokenPipeError:
            break

@tonio.main
def main():
    def close(sender, barrier):
        yield barrier.wait()
        sender.close()

    sender, receiver = channel.unbounded()
    barrier = sync.Barrier(3)
    yield tonio.spawn(*[
        producer(sender, barrier, 100),
        producer(sender, barrier, 200),
        consumer(receiver),
        consumer(receiver),
        consumer(receiver),
        consumer(receiver),
        close(sender, barrier),
    ])

await syntax

import tonio.colored as tonio
from tonio.colored import sync
from tonio.colored.sync import channel

async def producer(sender, barrier, offset):
    for i in range(20):
        message = offset + i
        sender.send(message)
    await barrier.wait()

async def consumer(receiver):
    while True:
        try:
            message = await receiver.receive()
            print(message)
        except BrokenPipeError:
            break

@tonio.main
async def main():
    async def close(sender, barrier):
        await barrier.wait()
        sender.close()

    sender, receiver = channel.unbounded()
    barrier = sync.Barrier(3)
    await tonio.spawn(*[
        producer(sender, barrier, 100),
        producer(sender, barrier, 200),
        consumer(receiver),
        consumer(receiver),
        consumer(receiver),
        consumer(receiver),
        close(sender, barrier),
    ])

Warning: receive is not cancel-safe in unbounded channels.

Non-blocking operations

Receivers of both channel kinds offer receive_nowait, a synchronous variant that never suspends: it returns the message or one of the Empty and Closed sentinels. Bounded senders, the only suspending ones, offer send_nowait in the same way: it returns None on success or one of the Full and Closed sentinels. The sentinels are available as attributes on the objects exposing them:

sender, receiver = channel.channel(8)

if sender.send_nowait(message) is sender.Full:
    ...
if (message := receiver.receive_nowait()) is receiver.Empty:
    ...

The same objects also expose try_send and try_receive, which raise instead: WouldBlock when the channel is full or empty, BrokenPipeError when it's closed.

Markers

The mark module provides decorator shortcuts for spawn_blocking and Semaphore:

yield syntax

import tonio

# run through `spawn_blocking`
@tonio.mark.blocking
def read_file(path):
    ...

# no more than 2 running at the same time
@tonio.mark.max_concurrency(2)
def fetch(url):
    ...

# `blocking` + `max_concurrency`
@tonio.mark.cpu_bound(4)
def resize(image):
    ...

await syntax

import tonio.colored as tonio

# run through `spawn_blocking`
@tonio.mark.blocking
def read_file(path):
    ...

# no more than 2 running at the same time
@tonio.mark.max_concurrency(2)
async def fetch(url):
    ...

# `blocking` + `max_concurrency`
@tonio.mark.cpu_bound(4)
def resize(image):
    ...

Network module

Network primitives are exposed under the tonio.net module.

Streams

The high-level network primitives in TonIO are centered around the SocketStream and SocketListener objects.

The SocketListener object implements an accept coroutine which returns a SocketStream object.
The SocketStream object implements the send_all and receive_some coroutines to send and receive data, and a send_eof method to shutdown the sending side.
Both objects implement a close method to shutdown the underlying socket.

You can create and interact with the above objects using some high-level helpers in the net module, specifically:

open_tcp_stream: a coroutine to open a SocketStream connected to a TCP endpoint
open_unix_socket: a coroutine to open a SocketStream connected to a Unix socket
open_tcp_listeners: a coroutine to initialise SocketListener objects
open_unix_listener: a coroutine to initialise a SocketListener on a Unix socket path
serve_listeners: a coroutine to spawn SocketListener accept loops targeting a handler
serve_tcp: a coroutine that joins open_tcp_listeners and serve_listeners in one call
serve_unix: a coroutine that joins open_unix_listener and serve_listeners in one call

yield syntax

from tonio.net import open_tcp_stream, serve_tcp

def server():
    yield serve_tcp(
        server_handle, 
        host='127.0.0.1', 
        port=8000
    )

def server_handle(stream):
    # receive some data
    data = yield stream.receive_some()

def client():
    stream = yield open_tcp_stream(
        host='127.0.0.1', 
        port=8000
    )
    # send some data
    yield stream.send_all(b"message")

await syntax

from tonio.colored.net import open_tcp_stream, serve_tcp

async def server():
    await serve_tcp(
        server_handle, 
        host='127.0.0.1', 
        port=8000
    )

async def server_handle(stream):
    # receive some data
    data = await stream.receive_some()

async def client():
    stream = await open_tcp_stream(
        host='127.0.0.1', 
        port=8000
    )
    # send some data
    await stream.send_all(b"message")

Unix domain sockets use the same objects, with serve_unix and open_unix_socket:

yield syntax

from tonio.net import open_unix_socket, serve_unix

def server():
    yield serve_unix(
        server_handle, 
        '/tmp/app.sock', 
        mode=0o600
    )

def client():
    stream = yield open_unix_socket(
        '/tmp/app.sock'
    )
    yield stream.send_all(b"message")

await syntax

from tonio.colored.net import open_unix_socket, serve_unix

async def server():
    await serve_unix(
        server_handle, 
        '/tmp/app.sock', 
        mode=0o600
    )

async def client():
    stream = await open_unix_socket(
        '/tmp/app.sock'
    )
    await stream.send_all(b"message")

Readiness and non-blocking operations

SocketStream also exposes its readiness state, for code that wants to decide when to read or write rather than just block on it:

wait_readable(timeout=None) and wait_writable(timeout=None): coroutines that suspend until the socket is ready, returning False if the timeout (in seconds) expires first
receive_some_nowait(max_bytes=None): a synchronous receive, returning the NotReady sentinel (available as stream.NotReady) when no data is available
try_receive_some(max_bytes=None): same, but raising WouldBlock instead
watch_readable() and watch_writable(): context managers producing a watcher, whose waiter() method returns a Waiter (or None when ready) you can combine with others, and whose ready() method tells whether the socket is ready

yield syntax

def handler(stream, stop):
    while True:
        with stream.watch_readable() as watcher:
            if (waiter := watcher.waiter()) is not None:
                yield waiter | stop.wait()
            if stop.is_set():
                break
        data = stream.receive_some_nowait()
        if data is stream.NotReady:
            continue
        ...

await syntax

async def handler(stream, stop):
    while True:
        with stream.watch_readable() as watcher:
            if (waiter := watcher.waiter()) is not None:
                await (waiter | stop.wait())
            if stop.is_set():
                break
        data = stream.receive_some_nowait()
        if data is stream.NotReady:
            continue
        ...

TLS streams

TonIO implement TLS wrappers around the streaming APIs through primitives in the tonio.net.tls module.

TonIO provides the TLSStream and TLSListener object wrappers and the following high-level helpers:

open_tls_over_tcp_stream: a coroutine to open a TLSStream wrapping a TCP SocketStream
open_tls_over_tcp_listeners: a coroutine to initialise TLSListener objects
serve_tls_over_tcp: a coroutine that joins open_tls_over_tcp_listeners and serve_listeners in one call

Low-level sockets

The tonio.net.socket module provides TonIO's basic low-level networking API.
Generally, the API exposed by this module mirrors the standard library socket module.

TonIO socket objects are overall very similar to the standard library socket objects, with the main difference being that blocking methods become coroutines.

yield syntax

import tonio
from tonio.net import socket

def server():
    sock = socket.socket()
    with sock:
        yield sock.bind(('127.0.0.1', 8000))
        sock.listen()

        while True:
            client, _ = yield sock.accept()
            tonio.spawn(server_handle(client))

def server_handle(connection):
    with connection:
        # receive some data
        data = yield connection.recv(4096)

def client():
    sock = socket.socket()
    with sock:
        yield sock.connect(('127.0.0.1', 8000))
        yield sock.send(b"message")

await syntax

import tonio.colored as tonio
from tonio.colored.net import socket

async def server():
    sock = socket.socket()
    with sock:
        await sock.bind(('127.0.0.1', 8000))
        sock.listen()

        while True:
            client, _ = await sock.accept()
            tonio.spawn(server_handle(client))

async def server_handle(connection):
    with connection:
        # receive some data
        data = await connection.recv(4096)

async def client():
    sock = socket.socket()
    with sock:
        await sock.connect(('127.0.0.1', 8000))
        await sock.send(b"message")

Filesystem module

TonIO's fs module exposes async API for filesystem operations (that are run in the blocking thread-pool). It provides open, a Path class mirroring pathlib.Path, and wrap_file to adopt an already-open file object.

yield syntax

import tonio
import tonio.fs as fs

def main():
    f = yield fs.open('data.txt', 'w')
    yield f.write('hello')
    yield f.close()

    path = fs.Path('data.txt')
    if (yield path.exists()):
        print((yield path.read_text()))

    for entry in (yield fs.Path('.').iterdir()):
        print(entry.name)

await syntax

import tonio.colored as tonio
import tonio.colored.fs as fs

async def main():
    async with await fs.open('data.txt', 'w') as f:
        await f.write('hello')

    path = fs.Path('data.txt')
    if await path.exists():
        print(await path.read_text())

    for entry in await fs.Path('.').iterdir():
        print(entry.name)

Operations that touch the filesystem are asynchronous; everything else stays synchronous.

Note: methods returning several paths (iterdir, glob, rglob, walk) are resolved in a single hop and give back a list. Since walk is fully materialised, mutating its dirnames does not prune the traversal, unlike pathlib.Path.walk.

Reading a file line by line differs between the two flavours. The await syntax supports async for and async with, neither of which the yield syntax can express:

yield syntax

def read_lines(path):
    f = yield fs.open(path, 'r')
    try:
        while True:
            line = yield f.readline()
            if not line:
                break
            print(line)
    finally:
        yield f.close()

await syntax

async def read_lines(path):
    async with await fs.open(path, 'r') as f:
        async for line in f:
            print(line)

Subprocesses

TonIO exposes two coroutines to run child processes: run_process for the common "run it and collect the outcome" case, and open_process for interacting with a process while it runs. Both spawn the process on the blocking thread-pool.

run_process returns a subprocess.CompletedProcess and accepts:

stdin: bytes to feed to the child (defaults to b'', meaning "close stdin immediately"), or a file descriptor/subprocess constant
capture_stdout / capture_stderr: when true, the relevant stream is collected and available on the result
check: when true (the default), a non-zero exit code raises subprocess.CalledProcessError

Any other keyword argument is forwarded to subprocess.Popen.

yield syntax

import tonio

def main():
    result = yield tonio.run_process(
        ['echo', 'hello'],
        capture_stdout=True
    )
    print(result.returncode)
    print(result.stdout)

await syntax

import tonio.colored as tonio

async def main():
    result = await tonio.run_process(
        ['echo', 'hello'],
        capture_stdout=True
    )
    print(result.returncode)
    print(result.stdout)

open_process returns a Process object instead, giving access to the running child. Passing subprocess.PIPE for stdin, stdout or stderr exposes the corresponding pipe as a stream on the process object, implementing the same send_all and receive_some coroutines of network streams:

yield syntax

import subprocess
import tonio

def main():
    proc = yield tonio.open_process(
        ['cat'],
        stdin=subprocess.PIPE,
        stdout=subprocess.PIPE
    )
    yield proc.stdin.send_all(b'hello')
    proc.stdin.close()
    data = yield proc.stdout.receive_some()
    code = yield proc.wait()

await syntax

import subprocess
import tonio.colored as tonio

async def main():
    proc = await tonio.open_process(
        ['cat'],
        stdin=subprocess.PIPE,
        stdout=subprocess.PIPE
    )
    await proc.stdin.send_all(b'hello')
    proc.stdin.close()
    data = await proc.stdout.receive_some()
    code = await proc.wait()

The Process object exposes:

args and pid: the command and the process identifier
stdin, stdout, stderr: the piped streams, or None when not piped
stdio: a (stdin, stdout) tuple, when both are piped
returncode and poll(): the exit code, or None while the process is still running
wait: a coroutine waiting for the process to exit, returning its exit code
send_signal, terminate, kill: synchronous methods to signal the process

Note: unlike run_process, open_process does not reap the child for you: remember to wait on it — possibly after a kill — otherwise the child outlives your task.

Note: processes in TonIO only communicate over unbuffered byte streams: the universal_newlines, text, encoding, errors and bufsize options of subprocess are not supported.

Note: on Windows, due to the platform's lack of features, the subprocess readiness implementation falls back to the blocking thread-pool. Thus, waiting on a process or read/write operations on pipes can't be interrupted while blocked: cancellations take effect only once the OS call returns.

Driving your own I/O

The io module exposes the primitives TonIO's own sockets, pipes and processes are built on, so you can plug any file descriptor the platform poller understands into the runtime, with the same readiness model.

register puts a descriptor under the poller's watch, for both reading and writing, and returns a ScheduledIO object. The registration is edge-triggered and lasts until you close it. The descriptor itself is neither owned nor switched to non-blocking mode: that's up to you.

Readiness is consumed in user space through a small protocol: arm_r (or arm_w) returns None if the descriptor is known to be ready, otherwise a Waiter to suspend on. Once ready, you perform the actual system call. If it would block anyway, clear_r (or clear_w) drops the stale readiness, so the next arm_r suspends again.

yield syntax

import os
import tonio

def read_some(sched, fd, max_bytes=65536):
    while True:
        if (waiter := sched.arm_r()) is not None:
            yield waiter
            continue
        try:
            return os.read(fd, max_bytes)
        except BlockingIOError:
            sched.clear_r()

@tonio.main
def main():
    r, w = os.pipe()
    os.set_blocking(r, False)
    sched = tonio.io.register(r)
    os.write(w, b"hello")
    print((yield read_some(sched, r)))
    sched.close()

await syntax

import os
import tonio.colored as tonio

async def read_some(sched, fd, max_bytes=65536):
    while True:
        if (waiter := sched.arm_r()) is not None:
            await waiter
            continue
        try:
            return os.read(fd, max_bytes)
        except BlockingIOError:
            sched.clear_r()

@tonio.main
async def main():
    r, w = os.pipe()
    os.set_blocking(r, False)
    sched = tonio.io.register(r)
    os.write(w, b"hello")
    print(await read_some(sched, r))
    sched.close()

arm_r and arm_w accept a timeout in seconds. An expired waiter just resumes, so a further arm_* call tells whether the descriptor got ready in the meantime. Readiness also covers hang-ups and errors: you'll be woken up when the peer goes away, and the following system call will report it.

consume_r and consume_w drain a direction's readiness and tell whether it was set, for descriptors signalling through readiness alone.

Descriptor streams

FdStream wraps a pipe-like descriptor into the same stream interface of the network module, taking ownership of it. It provides the send_all and receive_some coroutines, fileno and close, the latter also invoked when leaving a with block. This is what the pipes of Process objects are.

yield syntax

import os
import tonio
from tonio.io import FdStream

@tonio.main
def main():
    r, w = os.pipe()
    with FdStream(r) as reader, FdStream(w) as writer:
        yield writer.send_all(b"hello")
        print((yield reader.receive_some()))

await syntax

import os
import tonio.colored as tonio
from tonio.colored.io import FdStream

@tonio.main
async def main():
    r, w = os.pipe()
    with FdStream(r) as reader, FdStream(w) as writer:
        await writer.send_all(b"hello")
        print(await reader.receive_some())

Concurrent operations on the same direction of a stream raise WouldBlock, and a broken pipe surfaces as ResourceBroken.

Note: on Windows FdStream falls back to the blocking thread-pool, with the same limitations described for subprocesses.

Signals

TonIO provides a context manager to catch signals.

The usage of such context manager requires to first configure the runtime to listen for such signals:

yield syntax

import signal
import tonio
from tonio.time import interval

def sig_handle():
    with tonio.signal_receiver(
        signal.SIGHUP, 
        signal.SIGUSR1
    ) as sigs:
        for ev in sigs:
            sig = yield ev
            if sig == signal.SIGHUP:
                ...

@tonio.main(
    signals=[signal.SIGHUP, signal.SIGUSR1]
)
def main():
    tonio.spawn(sig_handle())
    ticker = interval(1)
    while True:
        yield ticker.tick()

await syntax

import signal
import tonio.colored as tonio
from tonio.colored.time import interval

async def sig_handle():
    with tonio.signal_receiver(
        signal.SIGHUP, 
        signal.SIGUSR1
    ) as sigs:
        async for sig in sigs:
            if sig == signal.SIGHUP:
                ...

@tonio.main(
    signals=[signal.SIGHUP, signal.SIGUSR1]
)
async def main():
    tonio.spawn(sig_handle())
    ticker = interval(1)
    while True:
        await ticker.tick()

Exceptions

The tonio.exceptions module exposes:

CancelledError: raised inside a coroutine when it gets cancelled
TimeoutError: raised when a timed operation expires
WouldBlock: raised by the or_raise and try_* variants when the operation cannot complete immediately
ResourceBroken: raised by streams when the underlying transport is unusable

Note: CancelledError and TimeoutError derive from BaseException, so they pass through except Exception clauses.

Testing

TonIO ships with a pytest plugin, which runs tests marked with the tonio marker on the runtime:

yield syntax

import pytest
import tonio

@pytest.mark.tonio
def test_sleep():
    yield tonio.sleep(0.1)

await syntax

import pytest
import tonio.colored as tonio

@pytest.mark.tonio
async def test_sleep():
    await tonio.sleep(0.1)

The marker can also be applied at module level with pytestmark = pytest.mark.tonio, or at class level.

To avoid marking tests entirely, the plugin also provides an auto mode, in which every async test gets run on the TonIO runtime. Auto mode can be enabled with the tonio_mode option in the pytest configuration:

[tool.pytest.ini_options]
tonio_mode = 'auto'

Async fixtures

For asynchronous fixtures where setup and teardown are required, the two syntaxes differ.

`await` syntax

Async fixtures involved in TonIO tests run on the runtime. Coroutine fixtures simply return their value, while async generator fixtures can yield it and run teardown code after the yield:

import pytest
from tonio.colored.net import open_tcp_stream

@pytest.fixture
async def connection():
    stream = await open_tcp_stream(host='127.0.0.1', port=8000)
    yield stream
    stream.close()

`yield` syntax

Given non-colored TonIO coroutines are indistinguishable from standard pytest generator fixtures, the plugin never runs generator fixtures on the runtime. This is usually not a limitation: since yield syntax tests already run as coroutines, simple setup can just happen within the test itself:

import pytest
from tonio.net import open_tcp_stream

def _connect():
    stream = yield open_tcp_stream(host='127.0.0.1', port=8000)
    return stream

@pytest.mark.tonio
def test_conn():
    stream = yield _connect()
    yield stream.send_all(b'ping')

When an actual teardown is required, the plugin provides the tonio_run fixture, which runs the given coroutine on the runtime:

@pytest.fixture
def connection(tonio_run):
    stream = tonio_run(_connect())

    def _disconnect():
        yield stream.send_all(b'bye')
        stream.close()

    yield stream
    tonio_run(_disconnect())

Runtime configuration in pytest

The TonIO runtime is always initialised once per test session, with context enabled and 2 threads. Runtime options can be customized overriding the session-scoped tonio_runtime_options fixture, for example in conftest.py:

import pytest

@pytest.fixture(scope='session')
def tonio_runtime_options():
    return {'threads': 4}

The runtime object itself is available to tests and fixtures via the session-scoped tonio_runtime fixture.

Libraries built on TonIO

In addition to the patches provided by the TonIO-Monkey project, the following libraries target TonIO natively:

httpunk: a low-level async HTTP library
punkreq: a high-level async HTTP client
punkasgi: an ASGI server built on TonIO

License

TonIO is released under the BSD License.

async

multithreading

python

Project health

PyPI

tonio

Latest release

0.10.4 · Oct 2, 2026

Releases this year

Significant stargazers

Marcelo Trylesinski

3,335 followers · starred Dec 2025

David Brochart

399 followers · starred Jun 2026

Patrick Arminio

1,007 followers · starred Dec 2025

Jair Henrique

248 followers · starred Dec 2025

Similar repos

awestlake87/pyo3-asyncio

350

AndreLouisCaron/a-tale-of-event-loops

Notes describing my understanding of how modern Python event loops work

331

PyO3/pyo3-async-runtimes

PyO3-based bridges between Python and Rust async runtimes

198

patrick-kidger/tinyio

Ever used asyncio and wished you hadn't? A tiny (~400 lines) event loop for Python.

739

python-trio/trio

Trio – a friendly Python library for async concurrency and I/O

7,349

Dicklesworthstone/asupersync

Async runtime for Rust where correctness is structural: region-owned tasks, cancel-correct…

285

alex-sherman/unsync

Unsynchronize asyncio

901

dabeaz/curio

Good Curio!

4,120

gi0baro/tonio

A multi-threaded async runtime for Python

Python

168

240 commits

updated Oct 2, 2026

See the code

README

TonIO

TonIO is a multi-threaded async runtime for free-threaded Python, built in Rust on top of the mio crate, and inspired by tinyio, trio and tokio.

Warning: TonIO is currently a work in progress. The APIs are subject to breaking changes.

Note: TonIO is available on free-threaded Python only. Windows is supported on a best-effort basis.

TonIO supports both using yield and the more canonical async/await notations, with the latter being available as part of the tonio.colored module. Following code snippets show both the usages.

Warning: despite the fact TonIO supports async and await notations, it's not compatible with any asyncio object like futures and tasks. The TonIO-Monkey project provides patches for some popular asyncio packages.

In a nutshell

yield syntax

import tonio

def wait_and_add(x):
    yield tonio.sleep(1)
    return x + 1

@tonio.main
def main():
    parallel = tonio.spawn(
        wait_and_add(3), 
        wait_and_add(4)
    )
    six = yield wait_and_add(5)
    four, five = yield parallel
    return four, five, six

assert main() == (4, 5, 6)

await syntax

import tonio.colored as tonio

async def wait_and_add(x):
    await tonio.sleep(1)
    return x + 1

@tonio.main
async def main():
    parallel = tonio.spawn(
        wait_and_add(3), 
        wait_and_add(4)
    )
    six = await wait_and_add(5)
    four, five = await parallel
    return four, five, six

assert main() == (4, 5, 6)

Usage

Entrypoint

Every TonIO program consists of an entrypoint, which should be passed to the run method:

yield syntax

import tonio

def main():
    yield
    print("Hello world")

tonio.run(main())

await syntax

import tonio.colored as tonio

async def main():
    await tonio.yield_now()
    print("Hello world")

tonio.run(main())

TonIO also provides a main decorator, thus we can rewrite the previous example as:

yield syntax

import tonio

@tonio.main
def main():
    yield
    print("Hello world")

main()

await syntax

import tonio.colored as tonio

@tonio.main
async def main():
    await tonio.yield_now()
    print("Hello world")

main()

Note: as you can see the colored module provides the additional yield_now coroutine, a quick way to define a suspension point, given you cannot just yield as in the non-colored notation.

Note: both run and main can only be called once per program. To run the runtime multiple times in the same program, follow the section below.

Manually managing the runtime

TonIO also provides the runtime function, to manually manage the runtime lifecycle:

import tonio

def _run1():
    ...

async def _run2():
    ...

def main():
    runtime = tonio.runtime()
    runtime.run_until_complete(_run1())
    runtime.run_until_complete(_run2())

Runtime options

The run, main and runtime methods accept options, specifically:

option name	description	default
`context`	enable `contextvars` usage in coroutines	`False`
`signals`	list of signals to listen to
`threads`	Number of runtime threads	# of CPU cores
`blocking_threadpool_size`	Maximum number of blocking threads	128
`blocking_threadpool_idle_ttl`	Idle timeout for blocking threads (in seconds)	30

Events

The core object in TonIO is Event. It's basically a wrapper around an atomic boolean flag, initialised with False. Event provides the following methods:

is_set(): return the value of the flag
set(): set the flag to True
clear(): set the flag to False
wait(timeout=None): returns a coroutine you can yield on that unblocks when the flag is set to True or the timeout expires. Timeout is in seconds.

yield syntax

import tonio

@tonio.main
def main():
    event = tonio.Event()

    def setter():
        yield tonio.sleep(1)
        event.set()

    tonio.spawn(setter())
    yield event.wait()

await syntax

import tonio.colored as tonio

@tonio.main
async def main():
    event = tonio.Event()

    async def setter():
        await tonio.sleep(1)
        event.set()

    tonio.spawn(setter())
    await event.wait()

Waiters

yield syntax

import tonio

@tonio.main
def main():
    ready, stop = tonio.Event(), tonio.Event()
    yield ready.wait() | stop.wait()
    if stop.is_set():
        return

await syntax

import tonio.colored as tonio

@tonio.main
async def main():
    ready, stop = tonio.Event(), tonio.Event()
    await (ready.wait() | stop.wait())
    if stop.is_set():
        return

Result

Spawning tasks

TonIO provides the spawn method to schedule new coroutines onto the runtime:

yield syntax

import tonio

def doubv(v):
    yield
    return v * 2

@tonio.main
def main():
    parallel = tonio.spawn(doubv(2), doubv(3))
    v3 = yield doubv(4)
    v1, v2 = yield parallel
    print([v1, v2, v3])

await syntax

import tonio.colored as tonio

async def doubv(v):
    await tonio.yield_now()
    return v * 2

@tonio.main
async def main():
    parallel = tonio.spawn(doubv(2), doubv(3))
    v3 = await doubv(4)
    v1, v2 = await parallel
    print([v1, v2, v3])

Coroutines passed to spawn get scheduled onto the runtime immediately. Using yield or await on the return value of spawn just waits for the coroutines to complete and retrieve the results.

Blocking tasks

TonIO provides the spawn_blocking method to schedule blocking operations onto the runtime:

yield syntax

import tonio

def read_file(path):
    with open(path, "r") as f:
        return f.read()

@tonio.main
def main():
    file_data = yield tonio.spawn_blocking(
        read_file, 
        "sometext.txt"
    )

await syntax

import tonio.colored as tonio

def read_file(path):
    with open(path, "r") as f:
        return f.read()

@tonio.main
async def main():
    file_data = await tonio.spawn_blocking(
        read_file, 
        "sometext.txt"
    )

Running tasks from synchronous contexts

Warning: using block_on from within a coroutine might produce a runtime deadlock.

Map utilities

TonIO provides the map and map_blocking utilities to spawn the same operation with an iterable of parameters:

yield syntax

import tonio

accum = []

def task(no):
    yield tonio.sleep(0.5)
    accum.append(no * 2)

@tonio.main
def main():
    yield tonio.map(task, range(4))

await syntax

import tonio.colored as tonio

accum = []

async def task(no):
    await tonio.sleep(0.5)
    accum.append(no * 2)

@tonio.main
async def main():
    await tonio.map(task, range(4))

Completion-based iterators

TonIO provides the as_completed utility to iterate over task results based on completion order:

yield syntax

import tonio

def _sleep(v):
    yield tonio.sleep(v)
    return v

@tonio.main
def main():
    vals = []
    for task in tonio.as_completed(
        _sleep(0.5),
        _sleep(0.1),
        _sleep(0.3),
    ):
        vals.append(yield task)

await syntax

import tonio.colored as tonio

async def _sleep(v):
    await tonio.sleep(v)
    return v

@tonio.main
async def main():
    vals = []
    async for val in tonio.as_completed(
        _sleep(0.5),
        _sleep(0.1),
        _sleep(0.3),
    ):
        vals.append(val)

Scopes and cancellations

Note: cancellations in TonIO are optimistic and terminal; optimistic means that cancelling takes effect only at suspension points, terminal means that once a coroutine is cancelled every later suspension is cancelled as well.

Warning: some TonIO primitives are not cancel-safe; they are marked as such in their sections. An operation is considered cancel-safe when cancelling it leaves the objects it operates on untouched.

TonIO provides a scope context, that lets you cancel work spawned within it:

yield syntax

import tonio

def slow_push(target, sleep):
    yield tonio.sleep(sleep)
    target.append(True)

@tonio.main
def main():
    values = []
    with tonio.scope() as scope:
        scope.spawn(slow_push(values, 0.1))
        scope.spawn(slow_push(values, 2))
        yield tonio.sleep(0.2)
        scope.cancel()
    yield scope()
    assert len(values) == 1

await syntax

import tonio.colored as tonio

async def slow_push(target, sleep):
    await tonio.sleep(sleep)
    target.append(True)

@tonio.main
async def main():
    values = []
    async with tonio.scope() as scope:
        scope.spawn(slow_push(values, 0.1))
        scope.spawn(slow_push(values, 2))
        await tonio.sleep(0.2)
        scope.cancel()
    assert len(values) == 1

Note: as you can see, the colored version of scope doesn't require to be awaited, as it will yield when exiting the context.

Select first completing task

TonIO also provides a select utility to cancel remaining work on the first completing task:

yield syntax

import tonio

def slow_push(target, sleep):
    yield tonio.sleep(sleep)
    target.append(True)

@tonio.main
def main():
    values = []
    yield tonio.select(
        slow_push(values, 0.1),
        slow_push(values, 2)
    )
    assert len(values) == 1

await syntax

import tonio.colored as tonio

async def slow_push(target, sleep):
    await tonio.sleep(sleep)
    target.append(True)

@tonio.main
async def main():
    values = []
    await tonio.select(
        slow_push(values, 0.1),
        slow_push(values, 2)
    )
    assert len(values) == 1

Note: select also accepts waiters, so an Event.wait() can race against coroutines.

tonio.time.time(): a function returning the runtime's clock (in seconds, microsecond resolution)
tonio.time.sleep(delay): a coroutine you can yield on to sleep (delay is in seconds)
tonio.time.timeout(coro, timeout): a coroutine you can yield on returning a tuple (output, success). If the coroutine succeeds in the given time then the pair (output, True) is returned. Otherwise this will return (None, False).

Note: time.sleep is also exported to the main tonio module.

Note: all of the above functions are also present in tonio.colored.time module.

Scheduling work

TonIO provides the time.interval function to create interval objects you can yield on a scheduled basis:

yield syntax

import tonio
from tonio import time

def some_task():
    ...

def scheduler():
    interval = time.interval(1)
    while True:
        yield interval.tick()
        tonio.spawn(some_task())

@tonio.main
def main():
    tonio.spawn(scheduler())
    # do some other work

await syntax

import tonio.colored as tonio
from tonio.colored import time

async def some_task():
    ...

async def scheduler():
    interval = time.interval(1)
    while True:
        await interval.tick()
        tonio.spawn(some_task())

@tonio.main
async def main():
    tonio.spawn(scheduler())
    # do some other work

from tonio import time

# tick every 500ms, with the first tick happening in 5 seconds from now
interval = time.interval(0.5, time.time() + 5)

Warning: Interval.tick is not cancel-safe.

Synchronization primitives

Synchronization primitives are exposed in the tonio.sync module.

Lock

Implements a classic mutex, or a non-reentrant, single-owner lock for coroutines:

yield syntax

import tonio
from tonio import sync

@tonio.main
def main():
    # counter can't go above 1
    counter = 0

    def _count(lock):
        nonlocal counter
        with (yield lock()):
            counter += 1
            yield
            counter -= 1
    
    lock = sync.Lock()
    yield tonio.spawn(*[
        _count(lock)
        for _ in range(10)
    ])

await syntax

import tonio.colored as tonio
from tonio.colored import sync

@tonio.main
async def main():
    # counter can't go above 1
    counter = 0

    async def _count(lock):
        nonlocal counter
        async with lock:
            counter += 1
            await tonio.yield_now()
            counter -= 1
    
    lock = sync.Lock()
    await tonio.spawn(*[
        _count(lock)
        for _ in range(10)
    ])

The Lock object also implements an or_raise method, that will immediately fail when the lock cannot be acquired:

from tonio.exceptions import WouldBlock

try:
    with lock.or_raise():
        ...
except WouldBlock:
    ...

Semaphore

A semaphore for coroutines:

yield syntax

import tonio
from tonio import sync

@tonio.main
def main():
    # counter can't go above 2
    counter = 0

    def _count(semaphore):
        nonlocal counter
        with (yield semaphore()):
            counter += 1
            yield
            counter -= 1
    
    semaphore = sync.Semaphore(2)
    yield tonio.spawn(*[
        _count(semaphore)
        for _ in range(10)
    ])

await syntax

import tonio.colored as tonio
from tonio.colored import sync

@tonio.main
async def main():
    # counter can't go above 2
    counter = 0

    async def _count(semaphore):
        nonlocal counter
        async with semaphore:
            counter += 1
            await tonio.yield_now()
            counter -= 1
    
    semaphore = sync.Semaphore(2)
    await tonio.spawn(*[
        _count(semaphore)
        for _ in range(10)
    ])

As for locks, the Semaphore object also implements an or_raise method, that will immediately fail when the lock cannot be acquired:

from tonio.exceptions import WouldBlock

try:
    with semaphore.or_raise():
        ...
except WouldBlock:
    ...

The Semaphore object also implements a tokens method, that returns the number of available tokens.

Barrier

A barrier for coroutines:

yield syntax

import tonio
from tonio import sync

@tonio.main
def main():
    barrier = sync.Barrier(3)
    count = 0

    def _start_at_3():
        nonlocal count
        count += 1
        i = yield barrier.wait()
        assert count == 3
        return i

    yield tonio.spawn(*[
        _start_at_3()
        for _ in range(3)
    ])

await syntax

import tonio.colored as tonio
from tonio.colored import sync

@tonio.main
async def main():
    barrier = sync.Barrier(3)
    count = 0

    async def _start_at_3():
        nonlocal count
        count += 1
        i = await barrier.wait()
        assert count == 3
        return i

    await tonio.spawn(*[
        _start_at_3()
        for _ in range(3)
    ])

Warning: Barrier.wait is not cancel-safe.

The Barrier object also implements a value method, which returns the current value of the barrier.

Channels

Multi-producer multi-consumer channels for inter-coroutine communication.

Bounded channel

yield syntax

import tonio
from tonio import sync
from tonio.sync import channel

def producer(sender, barrier, offset):
    for i in range(20):
        message = offset + i
        yield sender.send(message)
    yield barrier.wait()

def consumer(receiver):
    while True:
        try:
            message = yield receiver.receive()
            print(message)
        except BrokenPipeError:
            break

@tonio.main
def main():
    def close(sender, barrier):
        yield barrier.wait()
        sender.close()

    sender, receiver = channel.channel(2)
    barrier = sync.Barrier(3)
    yield tonio.spawn(*[
        producer(sender, barrier, 100),
        producer(sender, barrier, 200),
        consumer(receiver),
        consumer(receiver),
        consumer(receiver),
        consumer(receiver),
        close(sender, barrier),
    ])

await syntax

import tonio.colored as tonio
from tonio.colored import sync
from tonio.colored.sync import channel

async def producer(sender, barrier, offset):
    for i in range(20):
        message = offset + i
        await sender.send(message)
    await barrier.wait()

async def consumer(receiver):
    while True:
        try:
            message = await receiver.receive()
            print(message)
        except BrokenPipeError:
            break

@tonio.main
async def main():
    async def close(sender, barrier):
        await barrier.wait()
        sender.close()

    sender, receiver = channel.channel(2)
    barrier = sync.Barrier(3)
    await tonio.spawn(*[
        producer(sender, barrier, 100),
        producer(sender, barrier, 200),
        consumer(receiver),
        consumer(receiver),
        consumer(receiver),
        consumer(receiver),
        close(sender, barrier),
    ])

Warning: receive and send operations are not cancel-safe in bounded channels.

Unbounded channel

yield syntax

import tonio
from tonio import sync
from tonio.sync import channel

def producer(sender, barrier, offset):
    for i in range(20):
        message = offset + i
        sender.send(message)
    yield barrier.wait()

def consumer(receiver):
    while True:
        try:
            message = yield receiver.receive()
            print(message)
        except BrokenPipeError:
            break

@tonio.main
def main():
    def close(sender, barrier):
        yield barrier.wait()
        sender.close()

    sender, receiver = channel.unbounded()
    barrier = sync.Barrier(3)
    yield tonio.spawn(*[
        producer(sender, barrier, 100),
        producer(sender, barrier, 200),
        consumer(receiver),
        consumer(receiver),
        consumer(receiver),
        consumer(receiver),
        close(sender, barrier),
    ])

await syntax

import tonio.colored as tonio
from tonio.colored import sync
from tonio.colored.sync import channel

async def producer(sender, barrier, offset):
    for i in range(20):
        message = offset + i
        sender.send(message)
    await barrier.wait()

async def consumer(receiver):
    while True:
        try:
            message = await receiver.receive()
            print(message)
        except BrokenPipeError:
            break

@tonio.main
async def main():
    async def close(sender, barrier):
        await barrier.wait()
        sender.close()

    sender, receiver = channel.unbounded()
    barrier = sync.Barrier(3)
    await tonio.spawn(*[
        producer(sender, barrier, 100),
        producer(sender, barrier, 200),
        consumer(receiver),
        consumer(receiver),
        consumer(receiver),
        consumer(receiver),
        close(sender, barrier),
    ])

Warning: receive is not cancel-safe in unbounded channels.

Non-blocking operations

sender, receiver = channel.channel(8)

if sender.send_nowait(message) is sender.Full:
    ...
if (message := receiver.receive_nowait()) is receiver.Empty:
    ...

The same objects also expose try_send and try_receive, which raise instead: WouldBlock when the channel is full or empty, BrokenPipeError when it's closed.

Markers

The mark module provides decorator shortcuts for spawn_blocking and Semaphore:

yield syntax

import tonio

# run through `spawn_blocking`
@tonio.mark.blocking
def read_file(path):
    ...

# no more than 2 running at the same time
@tonio.mark.max_concurrency(2)
def fetch(url):
    ...

# `blocking` + `max_concurrency`
@tonio.mark.cpu_bound(4)
def resize(image):
    ...

await syntax

import tonio.colored as tonio

# run through `spawn_blocking`
@tonio.mark.blocking
def read_file(path):
    ...

# no more than 2 running at the same time
@tonio.mark.max_concurrency(2)
async def fetch(url):
    ...

# `blocking` + `max_concurrency`
@tonio.mark.cpu_bound(4)
def resize(image):
    ...

Network module

Network primitives are exposed under the tonio.net module.

Streams

The high-level network primitives in TonIO are centered around the SocketStream and SocketListener objects.

You can create and interact with the above objects using some high-level helpers in the net module, specifically:

open_tcp_stream: a coroutine to open a SocketStream connected to a TCP endpoint
open_unix_socket: a coroutine to open a SocketStream connected to a Unix socket
open_tcp_listeners: a coroutine to initialise SocketListener objects
open_unix_listener: a coroutine to initialise a SocketListener on a Unix socket path
serve_listeners: a coroutine to spawn SocketListener accept loops targeting a handler
serve_tcp: a coroutine that joins open_tcp_listeners and serve_listeners in one call
serve_unix: a coroutine that joins open_unix_listener and serve_listeners in one call

yield syntax

from tonio.net import open_tcp_stream, serve_tcp

def server():
    yield serve_tcp(
        server_handle, 
        host='127.0.0.1', 
        port=8000
    )

def server_handle(stream):
    # receive some data
    data = yield stream.receive_some()

def client():
    stream = yield open_tcp_stream(
        host='127.0.0.1', 
        port=8000
    )
    # send some data
    yield stream.send_all(b"message")

await syntax

from tonio.colored.net import open_tcp_stream, serve_tcp

async def server():
    await serve_tcp(
        server_handle, 
        host='127.0.0.1', 
        port=8000
    )

async def server_handle(stream):
    # receive some data
    data = await stream.receive_some()

async def client():
    stream = await open_tcp_stream(
        host='127.0.0.1', 
        port=8000
    )
    # send some data
    await stream.send_all(b"message")

Unix domain sockets use the same objects, with serve_unix and open_unix_socket:

yield syntax

from tonio.net import open_unix_socket, serve_unix

def server():
    yield serve_unix(
        server_handle, 
        '/tmp/app.sock', 
        mode=0o600
    )

def client():
    stream = yield open_unix_socket(
        '/tmp/app.sock'
    )
    yield stream.send_all(b"message")

await syntax

from tonio.colored.net import open_unix_socket, serve_unix

async def server():
    await serve_unix(
        server_handle, 
        '/tmp/app.sock', 
        mode=0o600
    )

async def client():
    stream = await open_unix_socket(
        '/tmp/app.sock'
    )
    await stream.send_all(b"message")

Readiness and non-blocking operations

SocketStream also exposes its readiness state, for code that wants to decide when to read or write rather than just block on it:

wait_readable(timeout=None) and wait_writable(timeout=None): coroutines that suspend until the socket is ready, returning False if the timeout (in seconds) expires first
receive_some_nowait(max_bytes=None): a synchronous receive, returning the NotReady sentinel (available as stream.NotReady) when no data is available
try_receive_some(max_bytes=None): same, but raising WouldBlock instead
watch_readable() and watch_writable(): context managers producing a watcher, whose waiter() method returns a Waiter (or None when ready) you can combine with others, and whose ready() method tells whether the socket is ready

yield syntax

def handler(stream, stop):
    while True:
        with stream.watch_readable() as watcher:
            if (waiter := watcher.waiter()) is not None:
                yield waiter | stop.wait()
            if stop.is_set():
                break
        data = stream.receive_some_nowait()
        if data is stream.NotReady:
            continue
        ...

await syntax

async def handler(stream, stop):
    while True:
        with stream.watch_readable() as watcher:
            if (waiter := watcher.waiter()) is not None:
                await (waiter | stop.wait())
            if stop.is_set():
                break
        data = stream.receive_some_nowait()
        if data is stream.NotReady:
            continue
        ...

TLS streams

TonIO implement TLS wrappers around the streaming APIs through primitives in the tonio.net.tls module.

TonIO provides the TLSStream and TLSListener object wrappers and the following high-level helpers:

open_tls_over_tcp_stream: a coroutine to open a TLSStream wrapping a TCP SocketStream
open_tls_over_tcp_listeners: a coroutine to initialise TLSListener objects
serve_tls_over_tcp: a coroutine that joins open_tls_over_tcp_listeners and serve_listeners in one call

Low-level sockets

The tonio.net.socket module provides TonIO's basic low-level networking API.
Generally, the API exposed by this module mirrors the standard library socket module.

TonIO socket objects are overall very similar to the standard library socket objects, with the main difference being that blocking methods become coroutines.

yield syntax

import tonio
from tonio.net import socket

def server():
    sock = socket.socket()
    with sock:
        yield sock.bind(('127.0.0.1', 8000))
        sock.listen()

        while True:
            client, _ = yield sock.accept()
            tonio.spawn(server_handle(client))

def server_handle(connection):
    with connection:
        # receive some data
        data = yield connection.recv(4096)

def client():
    sock = socket.socket()
    with sock:
        yield sock.connect(('127.0.0.1', 8000))
        yield sock.send(b"message")

await syntax

import tonio.colored as tonio
from tonio.colored.net import socket

async def server():
    sock = socket.socket()
    with sock:
        await sock.bind(('127.0.0.1', 8000))
        sock.listen()

        while True:
            client, _ = await sock.accept()
            tonio.spawn(server_handle(client))

async def server_handle(connection):
    with connection:
        # receive some data
        data = await connection.recv(4096)

async def client():
    sock = socket.socket()
    with sock:
        await sock.connect(('127.0.0.1', 8000))
        await sock.send(b"message")

Filesystem module

yield syntax

import tonio
import tonio.fs as fs

def main():
    f = yield fs.open('data.txt', 'w')
    yield f.write('hello')
    yield f.close()

    path = fs.Path('data.txt')
    if (yield path.exists()):
        print((yield path.read_text()))

    for entry in (yield fs.Path('.').iterdir()):
        print(entry.name)

await syntax

import tonio.colored as tonio
import tonio.colored.fs as fs

async def main():
    async with await fs.open('data.txt', 'w') as f:
        await f.write('hello')

    path = fs.Path('data.txt')
    if await path.exists():
        print(await path.read_text())

    for entry in await fs.Path('.').iterdir():
        print(entry.name)

Operations that touch the filesystem are asynchronous; everything else stays synchronous.

Note: methods returning several paths (iterdir, glob, rglob, walk) are resolved in a single hop and give back a list. Since walk is fully materialised, mutating its dirnames does not prune the traversal, unlike pathlib.Path.walk.

Reading a file line by line differs between the two flavours. The await syntax supports async for and async with, neither of which the yield syntax can express:

yield syntax

def read_lines(path):
    f = yield fs.open(path, 'r')
    try:
        while True:
            line = yield f.readline()
            if not line:
                break
            print(line)
    finally:
        yield f.close()

await syntax

async def read_lines(path):
    async with await fs.open(path, 'r') as f:
        async for line in f:
            print(line)

Subprocesses

run_process returns a subprocess.CompletedProcess and accepts:

stdin: bytes to feed to the child (defaults to b'', meaning "close stdin immediately"), or a file descriptor/subprocess constant
capture_stdout / capture_stderr: when true, the relevant stream is collected and available on the result
check: when true (the default), a non-zero exit code raises subprocess.CalledProcessError

Any other keyword argument is forwarded to subprocess.Popen.

yield syntax

import tonio

def main():
    result = yield tonio.run_process(
        ['echo', 'hello'],
        capture_stdout=True
    )
    print(result.returncode)
    print(result.stdout)

await syntax

import tonio.colored as tonio

async def main():
    result = await tonio.run_process(
        ['echo', 'hello'],
        capture_stdout=True
    )
    print(result.returncode)
    print(result.stdout)

yield syntax

import subprocess
import tonio

def main():
    proc = yield tonio.open_process(
        ['cat'],
        stdin=subprocess.PIPE,
        stdout=subprocess.PIPE
    )
    yield proc.stdin.send_all(b'hello')
    proc.stdin.close()
    data = yield proc.stdout.receive_some()
    code = yield proc.wait()

await syntax

import subprocess
import tonio.colored as tonio

async def main():
    proc = await tonio.open_process(
        ['cat'],
        stdin=subprocess.PIPE,
        stdout=subprocess.PIPE
    )
    await proc.stdin.send_all(b'hello')
    proc.stdin.close()
    data = await proc.stdout.receive_some()
    code = await proc.wait()

The Process object exposes:

args and pid: the command and the process identifier
stdin, stdout, stderr: the piped streams, or None when not piped
stdio: a (stdin, stdout) tuple, when both are piped
returncode and poll(): the exit code, or None while the process is still running
wait: a coroutine waiting for the process to exit, returning its exit code
send_signal, terminate, kill: synchronous methods to signal the process

Note: unlike run_process, open_process does not reap the child for you: remember to wait on it — possibly after a kill — otherwise the child outlives your task.

Note: processes in TonIO only communicate over unbuffered byte streams: the universal_newlines, text, encoding, errors and bufsize options of subprocess are not supported.

Note: on Windows, due to the platform's lack of features, the subprocess readiness implementation falls back to the blocking thread-pool. Thus, waiting on a process or read/write operations on pipes can't be interrupted while blocked: cancellations take effect only once the OS call returns.

Driving your own I/O

yield syntax

import os
import tonio

def read_some(sched, fd, max_bytes=65536):
    while True:
        if (waiter := sched.arm_r()) is not None:
            yield waiter
            continue
        try:
            return os.read(fd, max_bytes)
        except BlockingIOError:
            sched.clear_r()

@tonio.main
def main():
    r, w = os.pipe()
    os.set_blocking(r, False)
    sched = tonio.io.register(r)
    os.write(w, b"hello")
    print((yield read_some(sched, r)))
    sched.close()

await syntax

import os
import tonio.colored as tonio

async def read_some(sched, fd, max_bytes=65536):
    while True:
        if (waiter := sched.arm_r()) is not None:
            await waiter
            continue
        try:
            return os.read(fd, max_bytes)
        except BlockingIOError:
            sched.clear_r()

@tonio.main
async def main():
    r, w = os.pipe()
    os.set_blocking(r, False)
    sched = tonio.io.register(r)
    os.write(w, b"hello")
    print(await read_some(sched, r))
    sched.close()

consume_r and consume_w drain a direction's readiness and tell whether it was set, for descriptors signalling through readiness alone.

Descriptor streams

yield syntax

import os
import tonio
from tonio.io import FdStream

@tonio.main
def main():
    r, w = os.pipe()
    with FdStream(r) as reader, FdStream(w) as writer:
        yield writer.send_all(b"hello")
        print((yield reader.receive_some()))

await syntax

import os
import tonio.colored as tonio
from tonio.colored.io import FdStream

@tonio.main
async def main():
    r, w = os.pipe()
    with FdStream(r) as reader, FdStream(w) as writer:
        await writer.send_all(b"hello")
        print(await reader.receive_some())

Concurrent operations on the same direction of a stream raise WouldBlock, and a broken pipe surfaces as ResourceBroken.

Note: on Windows FdStream falls back to the blocking thread-pool, with the same limitations described for subprocesses.

Signals

TonIO provides a context manager to catch signals.

The usage of such context manager requires to first configure the runtime to listen for such signals:

yield syntax

import signal
import tonio
from tonio.time import interval

def sig_handle():
    with tonio.signal_receiver(
        signal.SIGHUP, 
        signal.SIGUSR1
    ) as sigs:
        for ev in sigs:
            sig = yield ev
            if sig == signal.SIGHUP:
                ...

@tonio.main(
    signals=[signal.SIGHUP, signal.SIGUSR1]
)
def main():
    tonio.spawn(sig_handle())
    ticker = interval(1)
    while True:
        yield ticker.tick()

await syntax

import signal
import tonio.colored as tonio
from tonio.colored.time import interval

async def sig_handle():
    with tonio.signal_receiver(
        signal.SIGHUP, 
        signal.SIGUSR1
    ) as sigs:
        async for sig in sigs:
            if sig == signal.SIGHUP:
                ...

@tonio.main(
    signals=[signal.SIGHUP, signal.SIGUSR1]
)
async def main():
    tonio.spawn(sig_handle())
    ticker = interval(1)
    while True:
        await ticker.tick()

Exceptions

The tonio.exceptions module exposes:

CancelledError: raised inside a coroutine when it gets cancelled
TimeoutError: raised when a timed operation expires
WouldBlock: raised by the or_raise and try_* variants when the operation cannot complete immediately
ResourceBroken: raised by streams when the underlying transport is unusable

Note: CancelledError and TimeoutError derive from BaseException, so they pass through except Exception clauses.

Testing

TonIO ships with a pytest plugin, which runs tests marked with the tonio marker on the runtime:

yield syntax

import pytest
import tonio

@pytest.mark.tonio
def test_sleep():
    yield tonio.sleep(0.1)

await syntax

import pytest
import tonio.colored as tonio

@pytest.mark.tonio
async def test_sleep():
    await tonio.sleep(0.1)

The marker can also be applied at module level with pytestmark = pytest.mark.tonio, or at class level.

[tool.pytest.ini_options]
tonio_mode = 'auto'

Async fixtures

For asynchronous fixtures where setup and teardown are required, the two syntaxes differ.

`await` syntax

Async fixtures involved in TonIO tests run on the runtime. Coroutine fixtures simply return their value, while async generator fixtures can yield it and run teardown code after the yield:

import pytest
from tonio.colored.net import open_tcp_stream

@pytest.fixture
async def connection():
    stream = await open_tcp_stream(host='127.0.0.1', port=8000)
    yield stream
    stream.close()

`yield` syntax

import pytest
from tonio.net import open_tcp_stream

def _connect():
    stream = yield open_tcp_stream(host='127.0.0.1', port=8000)
    return stream

@pytest.mark.tonio
def test_conn():
    stream = yield _connect()
    yield stream.send_all(b'ping')

When an actual teardown is required, the plugin provides the tonio_run fixture, which runs the given coroutine on the runtime:

@pytest.fixture
def connection(tonio_run):
    stream = tonio_run(_connect())

    def _disconnect():
        yield stream.send_all(b'bye')
        stream.close()

    yield stream
    tonio_run(_disconnect())

Runtime configuration in pytest

import pytest

@pytest.fixture(scope='session')
def tonio_runtime_options():
    return {'threads': 4}

The runtime object itself is available to tests and fixtures via the session-scoped tonio_runtime fixture.

Libraries built on TonIO

In addition to the patches provided by the TonIO-Monkey project, the following libraries target TonIO natively:

httpunk: a low-level async HTTP library
punkreq: a high-level async HTTP client
punkasgi: an ASGI server built on TonIO

License

TonIO is released under the BSD License.

async

multithreading

python

Project health

PyPI

tonio

Latest release

0.10.4 · Oct 2, 2026

Releases this year

Significant stargazers

Marcelo Trylesinski

3,335 followers · starred Dec 2025

David Brochart

399 followers · starred Jun 2026

Patrick Arminio

1,007 followers · starred Dec 2025

Jair Henrique

248 followers · starred Dec 2025

Similar repos

awestlake87/pyo3-asyncio

350

AndreLouisCaron/a-tale-of-event-loops

Notes describing my understanding of how modern Python event loops work

331

PyO3/pyo3-async-runtimes

PyO3-based bridges between Python and Rust async runtimes

198

patrick-kidger/tinyio

Ever used asyncio and wished you hadn't? A tiny (~400 lines) event loop for Python.

739

python-trio/trio

Trio – a friendly Python library for async concurrency and I/O

7,349

Dicklesworthstone/asupersync

Async runtime for Rust where correctness is structural: region-owned tasks, cancel-correct…

285

alex-sherman/unsync

Unsynchronize asyncio

901

dabeaz/curio

Good Curio!

4,120

gi0baro/tonio

README

TonIO

In a nutshell

Usage

Entrypoint

Manually managing the runtime

Runtime options

Events

Waiters

Result

Spawning tasks

Blocking tasks

Running tasks from synchronous contexts

Map utilities

Completion-based iterators

Scopes and cancellations

Select first completing task

Time-related functions

Scheduling work

Synchronization primitives

Lock

Semaphore

Barrier

Channels

Bounded channel

Unbounded channel

Non-blocking operations

Markers

Network module

Streams

Readiness and non-blocking operations

TLS streams

Low-level sockets

Filesystem module

Subprocesses

Driving your own I/O

Descriptor streams

Signals

Exceptions

Testing

Async fixtures

await syntax

yield syntax

Runtime configuration in pytest

Libraries built on TonIO

License

Project health

Significant stargazers

Similar repos

gi0baro/tonio

README

TonIO

In a nutshell

Usage

Entrypoint

Manually managing the runtime

Runtime options

Events

Waiters

Result

Spawning tasks

Blocking tasks

Running tasks from synchronous contexts

Map utilities

Completion-based iterators

Scopes and cancellations

Select first completing task

Time-related functions

Scheduling work

Synchronization primitives

Lock

Semaphore

Barrier

Channels

Bounded channel

Unbounded channel

Non-blocking operations

Markers

Network module

`await` syntax

`yield` syntax

`await` syntax

`yield` syntax