Migrating from v1 to v2

v2 is a complete reboot of Telethon v1. Because a lot of the library has suffered radical changes, there are no plans to provide “bridge” methods emulating the old interface. Doing so would take a lot of extra time and energy, and it’s honestly not fun.

What this means is that your v1 code very likely won’t run in v2. Sorry. I hope you can use this opportunity to shake up your dusty code into a cleaner design, too.

The common theme in v2 could be described as “no bullshit”.

v1 had grown a lot of features. A lot of them did a lot of things, at all once, in slightly different ways. Semver allows additions, so v2 will start out smaller and grow in a controlled manner.

Custom types were a frankestein monster, combining both raw and manually-defined properties in hacky ways. Type hinting was an unmaintained disaster. Features such as file IDs, proxies and a lot of utilities were pretty much abandoned.

The several attempts at making v2 a reality over the years starting from the top did not work out. A bottom-up approach was needed. So a full rewrite was warranted.

TLSharp was Telethon’s seed. Telethon v0 was needed to learn Python at all. Telethon v1 was necessary to learn what was a good design, and what wasn’t. This inspired grammers, a Rust re-implementation with a thought-out design. Telethon v2 completes the loop by porting grammers back to Python, now built with years of experience in the Telegram protocol.

It turns out static type checking is a very good idea for long-running projects. So I strongly encourage you to use mypy when developing code with Telethon v2. I can guarantee you will into far less problems.

Without further ado, let’s take a look at the biggest changes. This list may not be exhaustive, but it should give you an idea on what to expect. If you feel like a major change is missing, please open an issue.

Complete project restructure

The public modules under the telethon now make actual sense.

  • The root telethon package contains the basics like the Client and RpcError.

  • telethon.types contains all the types, for your tpye-hinting needs.

  • telethon.events contains all the events.

  • telethon.events.filters contains all the event filters.

  • telethon.session contains the session storages, should you choose to build a custom one.

  • telethon.errors is no longer a module. It’s actually a factory object returning new error types on demand. This means you don’t need to wait for new library versions to be released to catch them.


Be sure to check the documentation for telethon.errors to learn about error changes. Notably, errors such as FloodWaitError no longer have a .seconds field. Instead, every value for every error type is always .value.

This was also a good opportunity to remove a lot of modules that were not supposed to public in their entirety: .crypto, .extensions, .network, .custom, .functions, .helpers, .hints, .password, .requestiter, .sync, .types, .utils.

TelegramClient renamed to Client

You can rename it with as during import if you want to use the old name.

Python allows using namespaces via packages and modules. Therefore, the full name telethon.Client already indicates it’s from telethon, so the old Telegram prefix was redundant.

No telethon.sync hack

You can no longer import telethon.sync to have most calls wrapped in asyncio.loop.run_until_complete() for you.

Raw API is now private

v2 aims to comply with Semantic Versioning. This is impossible because Telegram is a live service that can change things any time. But we can get pretty close.

In v1, minor version changes bumped Telegram’s layer. This technically violated semver, because they were part of a public module.

To allow for new layers to be added without the need for major releases, telethon._tl is instead private. Here’s the recommended way to import and use it now:

from telethon import _tl as tl

was_reset = await client(tl.functions.account.reset_wall_papers())

if isinstance(chat, tl.abcs.User):
    if isinstance(chat, tl.types.UserEmpty):
    # chat is tl.types.User

There are three modules (four, if you count core, which you probably should not use). Each of them can have an additional namespace (as seen above with account.).

  • tl.functions contains every TL definition treated as a function. The naming convention now follows Python’s, and are snake_case.

  • tl.abcs contains every abstract class, the “boxed” types from Telegram. You can use these for your type-hinting needs.

  • tl.types contains concrete instances, the “bare” types Telegram actually returns. You’ll probably use these with isinstance() a lot.

Most custom types will also have a private _raw attribute with the original value from Telegram.

Raw API has a reduced feature-set

The string representation is now on object.__repr__(), not object.__str__().

All types use __slots__ to save space. This means you can’t add extra fields to these at runtime unless you subclass.

The .stringify() methods on all TL types no longer exists. Instead, you can use a library like beauty-print.

The .to_dict() method on all TL types no longer exists. The same is true for .to_json(). Instead, you can use a library like json-pickle or write your own:

def to_dict(obj):
    if obj is None or isinstance(obj, (bool, int, bytes, str)): return obj
    if isinstance(obj, list): return [to_dict(x) for x in obj]
    if isinstance(obj, dict): return {k: to_dict(v) for k, v in obj.items()}
    return {slot: to_dict(getattr(obj, slot)) for slot in obj.__slots__}

Lesser-known methods such as TLObject.pretty_format, serialize_bytes, serialize_datetime and from_reader are also gone. The remaining methods are:

  • Serializable.constructor_id() class-method, to get the integer identifier of the corresponding type constructor.

  • Serializable.from_bytes() class-method, to convert serialized bytes back into the class.

  • object.__bytes__() instance-method, to serialize the instance into bytes the way Telegram expects.

Functions are no longer a class with attributes. They serialize the request immediately. This means you cannot create request instance and change it later. Consider using functools.partial() if you want to reuse parts of a request instead.

Functions no longer have an asynchronous .resolve(). This used to let you pass usernames and have them be resolved to InputPeer automatically (unless it was nested).

Changes to start and client context-manager

You can no longer start() the client.

Instead, you will need to first connect() and then start the interactive_login().

In v1, the when using the client as a context-manager, start() was called. Since that method no longer exists, it now instead only connect() and disconnect().

This means you won’t get annoying prompts in your terminal if the session was not authorized. It also means you can now use the context manager even with custom login flows.

The old sign_in() method also sent the code, which was rather confusing. Instead, you must now request_login_code() as a separate operation.

The old log_out() was also renamed to sign_out() for consistency with sign_in().

The old is_user_authorized() was renamed to is_authorized() since it works for bot accounts too.

Unified client iter and get methods

The client no longer has client.iter_... methods.

Instead, the return a type that supports both await and async for:

messages = await client.get_messages(chat, 100)
# or
async for message in client.get_messages(chat, 100):


Client.get_messages() no longer has funny rules for the limit either. If you await it without limit, it will probably take a long time to complete. This is in contrast to v1, where get defaulted to 1 message and iter to no limit.

Removed client methods and properties

No client.parse_mode property.

Instead, you need to specify how the message text should be interpreted every time. In send_message(), use text=, markdown= or html=. In send_file() and friends, use one of the caption parameters.

No client.loop property.

Instead, you can use asyncio.get_running_loop().

No client.conversation() method.

Instead, you will need to design your own FSM. The simplest approach could be using a global states dictionary storing the next function to call:

from functools import partial

states = {}

async def conversation_entry_point(event):
    if fn := state.get(event.sender.id):
        await fn(event)
        await event.respond('Hi! What is your name?')
        state[event.sender.id] = handle_name

async def handle_name(event):
    await event.respond('What is your age?')
    states[event.sender.id] = partial(handle_age, name=event.text)

async def handle_age(event, name):
    age = event.text
    await event.respond(f'Hi {name}, I am {age} too!')
    del states[event.sender.id]

No client.kick_participant() method.

This is not a thing in Telegram. It was implemented by restricting and then removing the restriction.

The old client.edit_permissions() was renamed to Client.set_participant_restrictions(). This defines the restrictions a banned participant has applied (bans them from doing those things). Revoking the right to view messages will kick them. This rename should avoid confusion, as it is now clear this is not to promote users to admin status.

For administrators, client.edit_admin was renamed to Client.set_participant_admin_rights() for consistency.

You can also use the aliases on the Participant, types.Participant.set_restrictions() and types.Participant.set_admin_rights().

Note that a new method, Client.set_chat_default_restrictions(), must now be used to set a chat’s default rights. You can also use the alias on the Group, types.Group.set_default_restrictions().

No client.download_profile_photo() method.

You can simply use Client.download() now. Note that download() no longer supports downloading contacts as .vcard.

No client.set_proxy() method.

Proxy support is no longer built-in. They were never officially maintained. This doesn’t mean you can’t use them. You’re now free to choose your own proxy library and pass a different connector to the Client constructor.

This should hopefully make it clear that most connection issues when using proxies do not come from Telethon.

No client.set_receive_updates method.

It was not working as expected.

No client.catch_up() method.

You can still configure it when creating the Client, which was the only way to make it work anyway.

No client.action() method.

No client.takeout() method.

No client.qr_login() method.

No client.edit_2fa() method.

No client.get_stats() method.

No client.edit_folder() method.

No client.build_reply_markup() method.

No client.list_event_handlers() method.

These are out of scope for the time being. They might be re-introduced in the future if there is a burning need for them and are not difficult to maintain. This doesn’t mean you can’t do these things anymore though, since the Raw API is still available.

Telethon v2 is committed to not exposing the raw API under any public API of the telethon package. This means any method returning data from Telegram must have a custom wrapper object and be maintained too. Because the standards are higher, the barrier of entry for new additions and features is higher too.

Removed or renamed message properties and methods

Messages no longer have raw_text or message properties.

Instead, you can access the types.Message.text, text_markdown or text_html. These names aim to be consistent with caption_markdown and caption_html.

In v1, messages coming from a client used that client’s parse mode as some sort of “global state”. Based on the client’s parse mode, v1 message.text property would return different things. But not all messages did this! Those coming from the raw API had no client, so text couldn’t know how to format the message.

Overall, the old design made the parse mode be pretty hidden. This was not very intuitive and also made it very awkward to combine multiple parse modes.

The forward property is now forward_info. The forward_to method is now simply forward(). This makes it more consistent with the rest of message methods.

The is_reply, reply_to_msg_id and reply_to properties are now replied_message_id. The get_reply_message method is now get_replied_message(). This should make it clear that you are not getting a reply to the current message, but rather the message it replied to.

The to_id, via_input_bot, action_entities, button_count properties are also gone. Some were kept for backwards-compatibility, some were redundant.

The click method no longer exists in the message. Instead, find the right buttons to click on.

The download method no longer exists in the message. Instead, use download on the message’s file.

HMMMM WEB_PREVIEW VS LINK_PREVIEW… probs use link. we’re previewing a link not the web

Event and filters are now separate

Event types are no longer callable and do not have filters inside them. There is no longer nested class Event inside them either.

Instead, the event type itself is what the handler will actually be called with. Because filters are separate, there is no longer a need for v1 @events.register either. It also means you can combine filters with &, | and ~.

Filters are now normal functions that work with any event. Of course, this doesn’t mean all filters make sense for all events. But you can use them in an unified manner.

Filters no longer support asynchronous operations, which removes a footgun. This was most commonly experienced when using usernames as the chats filter in v1, and getting flood errors you couldn’t handle. In v2, you must pass a list of identifiers. This means getting those identifiers is up to you, and you can handle it in a way that is appropriated for your application.

See also

In-depth explanation for Updates.

Behaviour changes in events

Events produced by the client itself will now also be received as updates. This means, for example, that your events.NewMessage handlers will run when you use Client.send_message(). This is needed to properly handle updates.

In v1, there was a backwards-compatibility hack that flagged results from the client as their “own”. But in some rare cases, it was possible to still receive messages sent by the client itself in v1. The hack has been removed so now the library will consistently deliver all updates.

events.StopPropagation no longer exists. In v1, all handlers were always called. Now handlers are called in order until the filter for one returns True. The default behaviour is that handlers after that one are not called. This behaviour can be changed with the check_all_handlers flag in Client constructor.

events.CallbackQuery has been renamed to events.ButtonCallback and no longer also handles “inline bot callback queries”. This was a hacky workaround.

events.MessageRead no longer triggers when the contents of a message are read, such as voice notes being played.

Albums in Telegram are an illusion. There is no “album media”. There is only separate messages pretending to be a single message.

events.Album was a hack that waited for a small amount of time to group messages sharing the same grouped identifier. If you want to wait for a full album, you will need to wait yourself:

pending_albums = {}  # global for simplicity
async def gather_album(event, handler):
    if pending := pending_albums.get(event.grouped_id):
        pending_albums[event.grouped_id] = [event]
        # Wait for other events to come in. Adjust delay to your needs.
        # This will NOT work if sequential updates are enabled (spawn a task to do the rest instead).
        await asyncio.sleep(1)
        events = pending_albums.pop(grouped_id, [])
        await handler(events)

async def handler(event):
    if event.grouped_id:
        await gather_album(event, handle_album)
        await handle_message(event)

async def handle_album(events):
    ...  # do stuff with events

async def handle_message(event):
    ...  # do stuff with event

Note that the above code is not foolproof and will not handle more than one client. It might be possible for album events to be delayed for more than a second.

Note that messages that do not belong to an album can be received in-between an album.

Overall, it’s probably better if you treat albums for what they really are: separate messages sharing a grouped_id.

Streamlined chat, input_chat and chat_id

The same goes for sender, input_sender and sender_id. And also for get_chat, get_input_chat, get_sender and get_input_sender. Yeah, it was bad.

Instead, events with chat information now always have a .chat, with at least the .id. The same is true for the .sender, as long as the event has one with at least the user identifier.

This doesn’t mean the .chat or .sender will have all the information. Telegram may still choose to send their min version with only basic details. But it means you don’t have to remember 5 different ways of using chats.

To replace the concept of “input chats”, v2 introduces types.PeerRef. A “peer” represents either a User, Group or Channel, much like Telegram’s Peer. A “peer reference” represents just enough information to reference that peer without relying on Telethon’s cache. This is the most efficient way to call methods like Client.send_message() too.

The concept of “marked IDs” also no longer exists. This means v2 no longer supports the - or -100 prefixes on identifiers. Using the raw Peer to wrap the identifiers is gone, too. Instead, you’re strongly encouraged to use types.PeerRef instances.

The concepts of of “entity” or “peer” are unified to peer. Overall, dealing with users, groups and channels should feel a lot more natural.

See also

In-depth explanation for Peers, users and chats.

Other methods like client.get_peer_id, client.get_input_entity and client.get_entity are gone too. While not directly related, client.is_bot is gone as well. You can use Client.get_me() or read it from the session instead.

The telethon.utils package is gone entirely, so methods like utils.resolve_id no longer exist either.

Session cache no longer exists

At least, not the way it did before.

The v1 cache that allowed you to use just chat identifiers to call methods is no longer saved to disk.

Sessions now only contain crucial information to have a working client. This includes the server address, authorization key, update state, and some very basic details.

To work around this, you can use types.PeerRef, which is designed to be easy to store. This means your application can choose the best way to deal with them rather than being forced into Telethon’s session.

See also

In-depth explanation for Sessions.

StringSession no longer exists

If you need to serialize the session data to a string, you can use something like jsonpickle. Or even the built-in pickle followed by base64 or just bytes.hex(). But be aware that these approaches probably will not be compatible with additions to the Session.