34 KiB



Bugs and suggestions should be reported at: yt-dlp/yt-dlp/issues. Unless you were prompted to or there is another pertinent reason (e.g. GitHub fails to accept the bug report), please do not send bug reports via personal email. For discussions, join us in our discord server.

Please include the full output of yt-dlp when run with -vU, i.e. add -vU flag to your command line, copy the whole output and post it in the issue body wrapped in ``` for better formatting. It should look similar to this:

$ yt-dlp -vU <your command line>
[debug] Command-line config: ['-v', '']
[debug] Encodings: locale UTF-8, fs utf-8, out utf-8, pref UTF-8
[debug] yt-dlp version 2021.09.25 (zip)
[debug] Python version 3.8.10 (CPython 64bit) - Linux-5.4.0-74-generic-x86_64-with-glibc2.29
[debug] exe versions: ffmpeg 4.2.4, ffprobe 4.2.4
[debug] Proxy map: {}
Current Build Hash 25cc412d1d3c0725a1f2f5b7e4682f6fb40e6d15f7024e96f7afd572e9919535
yt-dlp is up to date (2021.09.25)

Do not post screenshots of verbose logs; only plain text is acceptable.

The output (including the first lines) contains important debugging information. Issues without the full output are often not reproducible and therefore will be closed as incomplete.

The templates provided for the Issues, should be completed and not removed, this helps aide the resolution of the issue.

Please re-read your issue once again to avoid a couple of common mistakes (you can and should use this as a checklist):

Is the description of the issue itself sufficient?

We often get issue reports that we cannot really decipher. While in most cases we eventually get the required information after asking back multiple times, this poses an unnecessary drain on our resources.

So please elaborate on what feature you are requesting, or what bug you want to be fixed. Make sure that it's obvious

  • What the problem is
  • How it could be fixed
  • How your proposed solution would look like

If your report is shorter than two lines, it is almost certainly missing some of these, which makes it hard for us to respond to it. We're often too polite to close the issue outright, but the missing info makes misinterpretation likely. We often get frustrated by these issues, since the only possible way for us to move forward on them is to ask for clarification over and over.

For bug reports, this means that your report should contain the complete output of yt-dlp when called with the -vU flag. The error message you get for (most) bugs even says so, but you would not believe how many of our bug reports do not contain this information.

If the error is ERROR: Unable to extract ... and you cannot reproduce it from multiple countries, add --write-pages and upload the .dump files you get somewhere.

Site support requests must contain an example URL. An example URL is a URL you might want to download, like There should be an obvious video present. Except under very special circumstances, the main page of a video service (e.g. is not an example URL.

Are you using the latest version?

Before reporting any issue, type yt-dlp -U. This should report that you're up-to-date. This goes for feature requests as well.

Is the issue already documented?

Make sure that someone has not already opened the issue you're trying to open. Search at the top of the window or browse the GitHub Issues of this repository. If there is an issue, subcribe to it to be notified when there is any progress. Unless you have something useful to add to the converation, please refrain from commenting.

Additionally, it is also helpful to see if the issue has already been documented in the youtube-dl issue tracker. If similar issues have already been reported in youtube-dl (but not in our issue tracker), links to them can be included in your issue report here.

Why are existing options not enough?

Before requesting a new feature, please have a quick peek at the list of supported options. Many feature requests are for features that actually exist already! Please, absolutely do show off your work in the issue report and detail how the existing similar options do not solve your problem.

Have you read and understood the changes, between youtube-dl and yt-dlp

There are many changes between youtube-dl and yt-dlp (changes to default behavior), and some of the options available have a different behaviour in yt-dlp, or have been removed all together (list of changes to options). Make sure you have read and understand the differences in the options and how this may impact your downloads before opening an issue.

Is there enough context in your bug report?

People want to solve problems, and often think they do us a favor by breaking down their larger problems (e.g. wanting to skip already downloaded files) to a specific request (e.g. requesting us to look whether the file exists before downloading the info page). However, what often happens is that they break down the problem into two steps: One simple, and one impossible (or extremely complicated one).

We are then presented with a very complicated request when the original problem could be solved far easier, e.g. by recording the downloaded video IDs in a separate file. To avoid this, you must include the greater context where it is non-obvious. In particular, every feature request that does not consist of adding support for a new site should contain a use case scenario that explains in what situation the missing feature would be useful.

Does the issue involve one problem, and one problem only?

Some of our users seem to think there is a limit of issues they can or should open. There is no limit of issues they can or should open. While it may seem appealing to be able to dump all your issues into one ticket, that means that someone who solves one of your issues cannot mark the issue as closed. Typically, reporting a bunch of issues leads to the ticket lingering since nobody wants to attack that behemoth, until someone mercifully splits the issue into multiple ones.

In particular, every site support request issue should only pertain to services at one site (generally under a common domain, but always using the same backend technology). Do not request support for vimeo user videos, White house podcasts, and Google Plus pages in the same issue. Also, make sure that you don't post bug reports alongside feature requests. As a rule of thumb, a feature request does not include outputs of yt-dlp that are not immediately related to the feature at hand. Do not post reports of a network error alongside the request for a new video service.

Is anyone going to need the feature?

Only post features that you (or an incapacitated friend you can personally talk to) require. Do not post features because they seem like a good idea. If they are really useful, they will be requested by someone who requires them.

Is your question about yt-dlp?

Some bug reports are completely unrelated to yt-dlp and relate to a different, or even the reporter's own, application. Please make sure that you are actually using yt-dlp. If you are using a UI for yt-dlp, report the bug to the maintainer of the actual application providing the UI. In general, if you are unable to provide the verbose log, you should not be opening the issue here.

If the issue is with youtube-dl (the upstream fork of yt-dlp) and not with yt-dlp, the issue should be raised in the youtube-dl project.

Are you willing to share account details if needed?

The maintainers and potential contributors of the project often do not have an account for the website you are asking support for. So any developer interested in solving your issue may ask you for account details. It is your personal discretion whether you are willing to share the account in order for the developer to try and solve your issue. However, if you are unwilling or unable to provide details, they obviously cannot work on the issue and it cannot be solved unless some developer who both has an account and is willing/able to contribute decides to solve it.

By sharing an account with anyone, you agree to bear all risks associated with it. The maintainers and yt-dlp can't be held responsible for any misuse of the credentials.

While these steps won't necessarily ensure that no misuse of the account takes place, these are still some good practices to follow.

  • Look for people with Member (maintainers of the project) or Contributor (people who have previously contributed code) tag on their messages.
  • Change the password before sharing the account to something random (use this if you don't have a random password generator).
  • Change the password after receiving the account back.

Is the website primarily used for piracy?

We follow youtube-dl's policy to not support services that is primarily used for infringing copyright. Additionally, it has been decided to not to support porn sites that specialize in fakes. We also cannot support any service that serves only DRM protected content.


Most users do not need to build yt-dlp and can download the builds or get them via the other installation methods.

To run yt-dlp as a developer, you don't need to build anything either. Simply execute

python -m yt_dlp

To run the test, simply invoke your favorite test runner, or execute a test file directly; any of the following work:

python -m unittest discover
python test/

See item 6 of new extractor tutorial for how to run extractor specific test cases.

If you want to create a build of yt-dlp yourself, you can follow the instructions here.

Adding new feature or making overarching changes

Before you start writing code for implementing a new feature, open an issue explaining your feature request and atleast one use case. This allows the maintainers to decide whether such a feature is desired for the project in the first place, and will provide an avenue to discuss some implementation details. If you open a pull request for a new feature without discussing with us first, do not be surprised when we ask for large changes to the code, or even reject it outright.

The same applies for changes to the documentation, code style, or overarching changes to the architecture

Adding support for a new site

If you want to add support for a new site, first of all make sure this site is not dedicated to copyright infringement. yt-dlp does not support such sites thus pull requests adding support for them will be rejected.

After you have ensured this site is distributing its content legally, you can follow this quick list (assuming your service is called yourextractor):

  1. Fork this repository

  2. Check out the source code with:

     git clone
  3. Start a new git branch with

     cd yt-dlp
     git checkout -b yourextractor
  4. Start with this simple template and save it to yt_dlp/extractor/

    from .common import InfoExtractor
    class YourExtractorIE(InfoExtractor):
        _VALID_URL = r'https?://(?:www\.)?yourextractor\.com/watch/(?P<id>[0-9]+)'
        _TESTS = [{
            'url': '',
            'md5': 'TODO: md5 sum of the first 10241 bytes of the video file (use --test)',
            'info_dict': {
                'id': '42',
                'ext': 'mp4',
                'title': 'Video title goes here',
                'thumbnail': r're:^https?://.*\.jpg$',
                # TODO more properties, either as:
                # * A value
                # * MD5 checksum; start the string with md5:
                # * A regular expression; start the string with re:
                # * Any Python type, e.g. int or float
        def _real_extract(self, url):
            video_id = self._match_id(url)
            webpage = self._download_webpage(url, video_id)
            # TODO more code goes here, for example ...
            title = self._html_search_regex(r'<h1>(.+?)</h1>', webpage, 'title')
            return {
                'id': video_id,
                'title': title,
                'description': self._og_search_description(webpage),
                'uploader': self._search_regex(r'<div[^>]+id="uploader"[^>]*>([^<]+)<', webpage, 'uploader', fatal=False),
                # TODO more properties (see yt_dlp/extractor/
  5. Add an import in yt_dlp/extractor/ Note that the class name must end with IE.

  6. Run python test/ TestDownload.test_YourExtractor (note that YourExtractor doesn't end with IE). This should fail at first, but you can continually re-run it until you're done. If you decide to add more than one test, the tests will then be named TestDownload.test_YourExtractor, TestDownload.test_YourExtractor_1, TestDownload.test_YourExtractor_2, etc. Note that tests with only_matching key in test's dict are not counted in. You can also run all the tests in one go with TestDownload.test_YourExtractor_all

  7. Make sure you have atleast one test for your extractor. Even if all videos covered by the extractor are expected to be inaccessible for automated testing, tests should still be added with a skip parameter indicating why the particular test is disabled from running.

  8. Have a look at yt_dlp/extractor/ for possible helper methods and a detailed description of what your extractor should and may return. Add tests and code for as many as you want.

  9. Make sure your code follows yt-dlp coding conventions and check the code with flake8:

     $ flake8 yt_dlp/extractor/
  10. Make sure your code works under all Python versions supported by yt-dlp, namely CPython and PyPy for Python 3.7 and above. Backward compatibility is not required for even older versions of Python.

  11. When the tests pass, add the new files, commit them and push the result, like this:

     $ git add yt_dlp/extractor/
     $ git add yt_dlp/extractor/
     $ git commit -m '[yourextractor] Add extractor'
     $ git push origin yourextractor
  12. Finally, create a pull request. We'll then review and merge it.

In any case, thank you very much for your contributions!

Tip: To test extractors that require login information, create a file test/local_parameters.json and add "usenetrc": true or your username and password in it:

    "username": "your user name",
    "password": "your password"

yt-dlp coding conventions

This section introduces a guide lines for writing idiomatic, robust and future-proof extractor code.

Extractors are very fragile by nature since they depend on the layout of the source data provided by 3rd party media hosters out of your control and this layout tends to change. As an extractor implementer your task is not only to write code that will extract media links and metadata correctly but also to minimize dependency on the source's layout and even to make the code foresee potential future changes and be ready for that. This is important because it will allow the extractor not to break on minor layout changes thus keeping old yt-dlp versions working. Even though this breakage issue may be easily fixed by a new version of yt-dlp, this could take some time, during which the extractor will remain broken.

Mandatory and optional metafields

For extraction to work yt-dlp relies on metadata your extractor extracts and provides to yt-dlp expressed by an information dictionary or simply info dict. Only the following meta fields in the info dict are considered mandatory for a successful extraction process by yt-dlp:

  • id (media identifier)
  • title (media title)
  • url (media download URL) or formats

The aforementioned metafields are the critical data that the extraction does not make any sense without and if any of them fail to be extracted then the extractor is considered completely broken. While all extractors must return a title, they must also allow it's extraction to be non-fatal.

For pornographic sites, appropriate age_limit must also be returned.

The extractor is allowed to return the info dict without url or formats in some special cases if it allows the user to extract usefull information with --ignore-no-formats-error - e.g. when the video is a live stream that has not started yet.

Any field apart from the aforementioned ones are considered optional. That means that extraction should be tolerant to situations when sources for these fields can potentially be unavailable (even if they are always available at the moment) and future-proof in order not to break the extraction of general purpose mandatory fields.


Say you have some source dictionary meta that you've fetched as JSON with HTTP request and it has a key summary:

meta = self._download_json(url, video_id)

Assume at this point meta's layout is:

    "summary": "some fancy summary text",
    "user": {
        "name": "uploader name"

Assume you want to extract summary and put it into the resulting info dict as description. Since description is an optional meta field you should be ready that this key may be missing from the meta dict, so that you should extract it like:

description = meta.get('summary')  # correct

and not like:

description = meta['summary']  # incorrect

The latter will break extraction process with KeyError if summary disappears from meta at some later time but with the former approach extraction will just go ahead with description set to None which is perfectly fine (remember None is equivalent to the absence of data).

If the data is nested, do not use .get chains, but instead make use of traverse_obj.

Considering the above meta again, assume you want to extract ["user"]["name"] and put it in the resulting info dict as uploader

uploader = traverse_obj(meta, ('user', 'name'))  # correct

and not like:

uploader = meta['user']['name']  # incorrect


uploader = meta.get('user', {}).get('name')  # incorrect


uploader = try_get(meta, lambda x: x['user']['name'])  # old utility

Similarly, you should pass fatal=False when extracting optional data from a webpage with _search_regex, _html_search_regex or similar methods, for instance:

description = self._search_regex(
    webpage, 'description', fatal=False)

With fatal set to False if _search_regex fails to extract description it will emit a warning and continue extraction.

You can also pass default=<some fallback value>, for example:

description = self._search_regex(
    webpage, 'description', default=None)

On failure this code will silently continue the extraction with description set to None. That is useful for metafields that may or may not be present.

Another thing to remember is not to try to iterate over None

Say you extracted a list of thumbnails into thumbnail_data and want to iterate over them

thumbnail_data = data.get('thumbnails') or []
thumbnails = [{
    'url': item['url'],
    'height': item.get('h'),
} for item in thumbnail_data if item.get('url')]  # correct

and not like:

thumbnail_data = data.get('thumbnails')
thumbnails = [{
    'url': item['url'],
    'height': item.get('h'),
} for item in thumbnail_data]  # incorrect

In this case, thumbnail_data will be None if the field was not found and this will cause the loop for item in thumbnail_data to raise a fatal error. Using or [] avoids this error and results in setting an empty list in thumbnails instead.

Alternately, this can be further simplified by using traverse_obj

thumbnails = [{
    'url': item['url'],
    'height': item.get('h'),
} for item in traverse_obj(data, ('thumbnails', lambda _, v: v['url']))]

or, even better,

thumbnails = traverse_obj(data, ('thumbnails', ..., {'url': 'url', 'height': 'h'}))

Provide fallbacks

When extracting metadata try to do so from multiple sources. For example if title is present in several places, try extracting from at least some of them. This makes it more future-proof in case some of the sources become unavailable.


Say meta from the previous example has a title and you are about to extract it like:

title = meta.get('title')

If title disappears from meta in future due to some changes on the hoster's side the title extraction would fail.

Assume that you have some another source you can extract title from, for example og:title HTML meta of a webpage. In this case you can provide a fallback like:

title = meta.get('title') or self._og_search_title(webpage)

This code will try to extract from meta first and if it fails it will try extracting og:title from a webpage, making the extractor more robust.

Regular expressions

Don't capture groups you don't use

Capturing group must be an indication that it's used somewhere in the code. Any group that is not used must be non capturing.


Don't capture id attribute name here since you can't use it for anything anyway.





Make regular expressions relaxed and flexible

When using regular expressions try to write them fuzzy, relaxed and flexible, skipping insignificant parts that are more likely to change, allowing both single and double quotes for quoted values and so on.


Say you need to extract title from the following HTML code:

<span style="position: absolute; left: 910px; width: 90px; float: right; z-index: 9999;" class="title">some fancy title</span>

The code for that task should look similar to:

title = self._search_regex(  # correct
    r'<span[^>]+class="title"[^>]*>([^<]+)', webpage, 'title')

which tolerates potential changes in the style attribute's value. Or even better:

title = self._search_regex(  # correct
    webpage, 'title', group='title')

which also handles both single quotes in addition to double quotes.

The code definitely should not look like:

title = self._search_regex(  # incorrect
    r'<span style="position: absolute; left: 910px; width: 90px; float: right; z-index: 9999;" class="title">(.*?)</span>',
    webpage, 'title', group='title')

or even

title = self._search_regex(  # incorrect
    r'<span style=".*?" class="title">(.*?)</span>',
    webpage, 'title', group='title')

Here the presence or absence of other attributes including style is irrelevant for the data we need, and so the regex must not depend on it

Keep the regular expressions as simple as possible, but no simpler

Since many extractors deal with unstructured data provided by websites, we will often need to use very complex regular expressions. You should try to use the simplest regex that can accomplish what you want. In other words, each part of the regex must have a reason for existing. If you can take out a symbol and the functionality does not change, the symbol should not be there.



_VALID_URL = r'https?://(?:www\.)?website\.com/(?:[^/]+/){3,4}(?P<display_id>[^/]+)_(?P<id>\d+)'


_VALID_URL = r'https?:\/\/(?:www\.)?website\.com\/[^\/]+/[^\/]+/[^\/]+(?:\/[^\/]+)?\/(?P<display_id>[^\/]+)_(?P<id>\d+)'

Do not misuse . and use the correct quantifiers (+*?)

Avoid creating regexes that over-match because of wrong use of quantifiers. Also try to avoid non-greedy matching (?) where possible since they could easily result in catastrophic backtracking


title = self._search_regex(r'<span\b[^>]+class="title"[^>]*>([^<]+)', webpage, 'title')


title = self._search_regex(r'<span\b.*class="title".*>(.+?)<', webpage, 'title')

Long lines policy

There is a soft limit to keep lines of code under 100 characters long. This means it should be respected if possible and if it does not make readability and code maintenance worse. Sometimes, it may be reasonable to go upto 120 characters and sometimes even 80 can be unreadable. Keep in mind that this is not a hard limit and is just one of many tools to make the code more readable.

For example, you should never split long string literals like URLs or some other often copied entities over multiple lines to fit this limit:

Conversely, don't unnecessarily split small lines further. As a rule of thumb, if removing the line split keeps the code under 80 characters, it should be a single line.







uploader = traverse_obj(info, ('uploader', 'name'), ('author', 'fullname'))


uploader = traverse_obj(
    ('uploader', 'name'),
    ('author', 'fullname'))


formats = self._extract_m3u8_formats(
    m3u8_url, video_id, 'mp4', 'm3u8_native', m3u8_id='hls',
    note='Downloading HD m3u8 information', errnote='Unable to download HD m3u8 information')


formats = self._extract_m3u8_formats(m3u8_url,
                                     note='Downloading HD m3u8 information',
                                     errnote='Unable to download HD m3u8 information')


Always use single quotes for strings (even if the string has ') and double quotes for docstrings. Use ''' only for multi-line strings. An exception can be made if a string has multiple single quotes in it and escaping makes it significantly harder to read. For f-strings, use you can use double quotes on the inside. But avoid f-strings that have too many quotes inside.

Inline values

Extracting variables is acceptable for reducing code duplication and improving readability of complex expressions. However, you should avoid extracting variables used only once and moving them to opposite parts of the extractor file, which makes reading the linear flow difficult.



return {
    'title': self._html_search_regex(r'<h1>([^<]+)</h1>', webpage, 'title'),
    # ...some lines of code...


TITLE_RE = r'<h1>([^<]+)</h1>'
# ...some lines of code...
title = self._html_search_regex(TITLE_RE, webpage, 'title')
# ...some lines of code...
return {
    'title': title,
    # ...some lines of code...

Collapse fallbacks

Multiple fallback values can quickly become unwieldy. Collapse multiple fallback values into a single expression via a list of patterns.



description = self._html_search_meta(
    ['og:description', 'description', 'twitter:description'],
    webpage, 'description', default=None)


description = (
    self._og_search_description(webpage, default=None)
    or self._html_search_meta('description', webpage, default=None)
    or self._html_search_meta('twitter:description', webpage, default=None))

Methods supporting list of patterns are: _search_regex, _html_search_regex, _og_search_property, _html_search_meta.

Trailing parentheses

Always move trailing parentheses used for grouping/functions after the last argument. On the other hand, multi-line literal list/tuple/dict/set should closed be in a new line. Generators and list/dict comprehensions may use either style



url = traverse_obj(info, (
    'context', 'dispatcher', 'stores', 'VideoTitlePageStore', 'data', 'video', 0, 'VideoUrlSet', 'VideoUrl'), list)


url = traverse_obj(
    ('context', 'dispatcher', 'stores', 'VideoTitlePageStore', 'data', 'video', 0, 'VideoUrlSet', 'VideoUrl'),


url = traverse_obj(
    ('context', 'dispatcher', 'stores', 'VideoTitlePageStore', 'data', 'video', 0, 'VideoUrlSet', 'VideoUrl'),


f = {
    'url': url,
    'format_id': format_id,


f = {'url': url,
     'format_id': format_id}


formats = [process_formats(f) for f in format_data
           if f.get('type') in ('hls', 'dash', 'direct') and f.get('downloadable')]


formats = [
    process_formats(f) for f in format_data
    if f.get('type') in ('hls', 'dash', 'direct') and f.get('downloadable')

Use convenience conversion and parsing functions

Wrap all extracted numeric data into safe functions from yt_dlp/ int_or_none, float_or_none. Use them for string to number conversions as well.

Use url_or_none for safe URL processing.

Use traverse_obj and try_call (superseeds dict_get and try_get) for safe metadata extraction from parsed JSON.

Use unified_strdate for uniform upload_date or any YYYYMMDD meta field extraction, unified_timestamp for uniform timestamp extraction, parse_filesize for filesize extraction, parse_count for count meta fields extraction, parse_resolution, parse_duration for duration extraction, parse_age_limit for age_limit extraction.

Explore yt_dlp/ for more useful convenience functions.


description = traverse_obj(response, ('result', 'video', 'summary'), expected_type=str)
thumbnails = traverse_obj(response, ('result', 'thumbnails', ..., 'url'), expected_type=url_or_none)
video = traverse_obj(response, ('result', 'video', 0), default={}, expected_type=dict)
duration = float_or_none(video.get('durationMs'), scale=1000)
view_count = int_or_none(video.get('views'))

My pull request is labeled pending-fixes

The pending-fixes label is added when there are changes requested to a PR. When the necessary changes are made, the label should be removed. However, despite our best efforts, it may sometimes happen that the maintainer did not see the changes or forgot to remove the label. If your PR is still marked as pending-fixes a few days after all requested changes have been made, feel free to ping the maintainer who labeled your issue and ask them to re-review and remove the label.


See for instructions on how to embed yt-dlp in another Python program