View on GitHub


Publish or capture VDO.Ninja streams with Python (Raspberry Pi, Linux, Mac, Windows WSL)

Raspberry Ninja - not just for Raspberry Pis!

Turn your Raspberry Pi, Nvidia Jetson, Orange Pi, Windows PC, Mac, Linux box, or nearly any Python-compatible system into a ninja-cam with hardware-acceleration enabled! This lets you publish live streaming video and audio directly to your web browser or OBS instance using VDO.Ninja. Achieve very low streaming latency over the Internet or a LAN; all for free.

It also has the ability to record remote VDO.Ninja streams to disk (no transcode step), it can broadcast a low-latency video stream to multiple viewers (with a built-in SFU) at time, and because it works with VDO.Ninja, you get access to its ecosystem and related features.

AV1 support is now available, however you will need a powerful system or one with an AV1 hardware encoder to make use of it. WHIP-output support has also been added.


The core concepts and code used in this project can be reused for other projects; most Linux systems, and a large variety of embedded systems; potentially even smartphones. There’s a focus of supporting Raspberry Pis and Nvidia Jetson systems, which includes offering pre-built images and install scripts. Other Linux system users should still be able to use the code, but setup support will be limited.

Youtube video demoing:

I also have another longer YouTube video here, which focuses on setting up the Raspberry_ninja for IRL-streaming.**


Recent updates to Raspberry Ninja have added improved error correction and video redundency, along with automated dynamic bitrate controls for congestion management. This has greatly improved stream reliabilty, reducing frame loss, and limiting buffer sizes. That said, having more than 5-megabites of upload bandwidth and having a solid connection is recommend if intending to use the default settings.

Install options

See below for different install options

Setup for a Raspberry Pi

See the raspberry_pi sub-folder for instructions on installing and setting up a Raspberry Pi. Jump there now

A Raspberry Pi works fairly well with a CSI-connected camera, but USB-based cameras currently struggle a bit with older Raspberry Pi models. As a result, consider buying an Nvidia Jetson Nano 2GB instead of a Raspberry Pi if looking to jump into this all. Also, the RPI Zero W 1 and RPi 3 both don’t have the greatest WiFi built-in, while the Raspberry Pi Zero 2 seems to work rather well. Without good connectivity, you may find yourself facing frame-drops and stutter. HDMI to CSI adapters do work, but they may be limited to 25-fps and can be finicky still with some camera sources; audio over HDMI is also a bit tricky to setup currently.


Setup for an Nvidia Jetson

Please see the nvidia_jetson folder for details on installation. Jump there now


Nvidia Jetsons work well with USB-connected cameras and have a selection of compatible CSI-cameras well. You may need to buy WiFi adapter if it is not included.

Setup for Linux Desktops

You can deploy Raspberry.Ninja to a desktop pretty quickly in most cases, without compiling anything. I have an installer for recent versions of Ubuntu if interested. Jump there now

For other distros, see below for requirements

Requirements for linux systems in general:

You’ll want to install Gstreamer 1.16 or newer; emphasis on the newer. You’ll need to ensure libnice, srtp, sctp, and webrtcbin are part of that install, along with any media codecs you intend to use.

Python3 is also required, along with websockets. If you have PIP installed, pip3 install websockets can get you going there.

Setup for Windows (WSL)

You can actually run Raspberry Ninja on a Windows PC via the WSL virtual machine interface. It’s really quick and simple, except getting camera/hardware support going is tricky.

Still, it might be useful if you want to pull a stream from a remote Raspberry.Ninja system, recording the stream to disk or using it for local machine learning.

See the WSL install script here: Jump there now

It is possible to install Gstreamer for Windows natively, but due to the difficultly in that all, I’m not supporting it officially at present. The main challenge is cairo fails to compile, so that needs to be fixed first.

Setup for Mac OS X

Raspberry.Ninja can even run on a Mac! Although it’s not as streamlined a process as it could be, I’d say the difficulty is 4/10.

See the Mac OS X install script here: Jump there now

Generic quick-install method

Many modern versions of Linux distributions, such as Ubuntu 22, support Raspberry.Ninja with minimal installation effort.

The basic install script for Ubuntu-like systems is as below:

sudo apt-get update && sudo apt upgrade -y

 # Use a virtual environment or delete the following file if having issues
sudo rm /usr/lib/python3.11/EXTERNALLY-MANAGED ## For Debian 12-based systems

sudo apt-get install python3-pip -y

sudo apt-get install -y libgstreamer1.0-dev libgstreamer-plugins-base1.0-dev libgstreamer-plugins-bad1.0-dev gstreamer1.0-plugins-base gstreamer1.0-plugins-good gstreamer1.0-plugins-bad gstreamer1.0-plugins-ugly gstreamer1.0-libav gstreamer1.0-tools gstreamer1.0-x python3-pyqt5 python3-opengl gstreamer1.0-alsa gstreamer1.0-gl gstreamer1.0-qt5 gstreamer1.0-gtk3 gstreamer1.0-pulseaudio gstreamer1.0-nice gstreamer1.0-plugins-base-apps

pip3 install websockets
pip3 install cryptography

sudo apt-get install -y libcairo-dev ## possibly optional
pip3 install PyGObject ## possibly optional

sudo apt-get install git -y
cd ~ 
git clone
cd raspberry_ninja
python3 --test

Package managers with old versions of Gstreamers, or with no hardware acceleration or limited codec support, may be limited in what Raspberry.Ninja can offer. For the most up-to-date and comprehensive feature set, compiling Gstreamer from scratch may be still needed.

If wanting to use AV1 streaming, you’ll need to install gst-plugins-rs as well (av1parse and av1pay are needed), plus whatever AV1 encoder you wish to use.


Major updates sometimes will require that the latest Rasbperry Pi or Jetson image be installed on your device, but most updates are minor and only require the file to be updated. If you’ve just installed the latest device image, you will still want to update before going further, as the image is not updated with every new code release.

You can normally update by logging into your device, either via SSH, or via mouse/keyboard with the terminal app open.

cd ~
cd raspberry_ninja
git pull

That’s it.

If you run into issues due making changes to the code, you can either git stash your changes first, or you can just delete the raspberry_ninja folder and clone it again.


cd ~
rm raspberry_ninja -r
git clone
cd raspberry_ninja

Updates are usually optional, as they typically just focus on added features or improving video quality/stability. I do recommend checking for new updates every now and then.


You should be able to run the publshing script simply with python3, however lots of options are available for customizing as desired.

$ python3

To get the list of supported commands with your version of the code, run python3 --help.

Sample help output: ( what’s shown below may not be up-to-date)

usage: [-h] [--streamid STREAMID] [--server SERVER]
                  [--bitrate BITRATE] [--width WIDTH] [--height HEIGHT]
                  [--framerate FRAMERATE] [--test] [--hdmi] [--v4l2 V4L2]
                  [--rpicam] [--nvidiacsi] [--alsa ALSA] [--pulse PULSE]
                  [--raw] [--h264] [--nvidia] [--rpi] [--novideo] [--noaudio]
                  [--pipeline PIPELINE]

optional arguments:
  -h, --help            show this help message and exit
  --streamid STREAMID   Stream ID of the peer to connect to
  --server SERVER       Handshake server to use, eg:
  --bitrate BITRATE     Sets the video bitrate. This is not adaptive, so
                        packet loss and insufficient bandwidth will cause
                        frame loss
  --width WIDTH         Sets the video width. Make sure that your input
                        supports it.
  --height HEIGHT       Sets the video height. Make sure that your input
                        supports it.
  --framerate FRAMERATE
                        Sets the video framerate. Make sure that your input
                        supports it.
  --test                Use test sources.
  --rtmp                Use RTMP instead of webRTC; pass "rtmp://"
  --whip                Use WHIP output instead of VDO.Ninja; pass ""
  --hdmi                Try to setup a HDMI dongle
  --camlink             Try to use Elgato Camlink as source
  --v4l2 V4L2           Sets the V4L2 input device.
  --rpicam              Sets the RaspberryPi input device.
  --nvidiacsi           Sets the input to the nvidia csi port.
  --alsa ALSA           Use alsa audio input.
  --pulse PULSE         Use pulse audio (or pipewire) input.
  --raw                 Opens the V4L2 device with raw capabilities.
  --bt601               Changes the input color profile when in raw mode to BT601
  --h264                For PC, instead of VP8, use x264.
  --vp8                 VP8 encoder instead of h264; likely software-based
  --aom                 Use AV1-AOM software encoding for publishing
  --av1                 Whatever AV1 encoder the script can find; more can be added on request
  --qsv                 Intel quicksync avi encoder
  --nvidia              Creates a pipeline optimized for Nvidia (Jetson?) hardware encoders
  --rpi                 Creates a pipeline optimized for Raspberry Pi (0,1,2,3,4) hardware encoders (⚠️RPi-5 doesn't have a hardware encoder)
  --novideo             Disables video input.
  --noaudio             Disables audio input.
  --omx                 An alternative hardware encoder for the RPi; glitches, but faster?
  --pipeline PIPELINE   A full custom pipeline
  --record STREAMID     Specify a remote stream ID to record; this will disable publishing mode
  --midi                MIDI transport; can forward/recieve MIDI to remote browser/device
  --save                Will save a local copy of the outbound stream to disk (MKV format)
  --rotate DEGREES      Will rotate the video by 90, 180 , or 270 degrees
  --multiviewer         Allows for multiple viewers at a time; this can increase bandwidth usage of course
  --nored               Disable error correction. If you don't disable it, the bandwidth may be up to 2x higher than the target video bitrate.  I do not recommend removing, unless you're on a pristine connection.
  --noqos               This will disable the qos feature. The QOS feature will lower the bitrate of the video encoder if heavy packet loss is detected. It won't lower it more than 5x (20% of target), but I find this works well to combat times where the network bandwidth is insufficient. 
  --pipein              Lets you pipe data in from a unix pipe, something like: `ffmpeg -i input.mp4 -o - | python3 --pipein auto`
  --libcamera           Use libcamera as a source; this may be needed if using third party cameras like those from Arducam

Changing video input sources

Using gst-device-monitor-1.0 will list available devices and their ‘caps’, or settings. This can help determine what GStreamer pipeline changes need to be made in the script or getting info about what video format options are available for your device.

To help further debug, gst-launch-1.0 can be used to test a pipeline out before adding it to the script. For for added reference, here is an example Pipeline for the Rasbperry Pi to enable UVC-based MJPEG video capture support is:

gst-launch-1.0 v4l2src device=/dev/video0 io-mode=2 ! image/jpeg,framerate=30/1,width=1920,height=1080 ! jpegparse ! nvjpegdec ! video/x-raw ! nvvidconv ! "video/x-raw(memory:NVMM)" ! omxh264enc ! "video/x-h264, stream-format=(string)byte-stream" ! h264parse ! rtph264pay config-interval=-1 ! application/x-rtp,media=video,encoding-name=H264,payload=96 ! fakesink

Notice how we used device = OUR_AUDIO_DEVICE_NAME to specify the audio device we want to use, and we configure the device to read and decode JPEG, as that is what our device in this case supports.

The Raspberry_Ninja script automatically tries to create a pipeline for you, based on the command line arguments passed, but you can override that at a code level with your own pipeline if easier as well.

Adding an audio source

The script will use the default system ALSA audio output device, although you can override that using the command line arguments or via manually setting a gstreamer pipeline at the code level.

To get details of available audio devices, assuming pulseaudio is installed, running the following from the command line will give us access to audio device IDs

 pactl list | grep -A2 'Source #' | grep 'Name: ' | cut -d" " -f2

resulting in..


In this example, an HDMI audio source is the first in the list, so that is our device name. Your device name will likely vary.

Pulse audio and ALSA audio command-line arguments can be passed to setup audio, without needing to tweak Gstreamer pipelines manually. The defaults I think will use the system ALSA default device.

How to Run:

Ensure the pi/jetson is connected to the Internet, via Ethernet is recommended for best performance. You’ll also very likely need to ensure a camera and/or microphone input are connected; this can also be a USB UVC device, supported CSI-based camera, or other selectable media inputs. It technically might be possible to even select a pipe to stream from, although this is a fairly advanced option.

Run using: python3 --streamid SomeStreamID --bitrate 2000

In Chrome, open this link to view:

You can have multiple viewers at a time, but you must enable that with a command-line argument.

Also note, if you run with sudo, you might get a permissions error when using audio.

Auto-starting the script on boot

A guide on how to setup a RPI to auto-stream on boot can be found in the Rasbperry Pi folder, along with details on how to configure the WiFi SSID and password without needing to SSH in first.

RTMP output

RTMP support overrides WebRTC support at the moment, and the features that are support are pretty limited.

python3 --rtmp rtmp:// --bitrate 6000

Things like bitrate, width, height, raw, framerate are also supported, but not a whole lot else.

RTMP support is currently experimental; example use with a Jetson here:

You can’t publish to with RTMP, but rather a service like YouTube.

SRT support

I have added SRT support to the Raspberry Pi image. You need to use it via Ffmpeg or Gstreamer via command line currently, as I haven’t added it to the Raspberry Ninja code directly yet. Still, it’s easy enough to publish via command line with SRT, and you get the benefits of an up-to-date Raspberry Pi image with drivers and software all pre-installed.

WHIP / Meshcast support

I added WHIP/WHEP dependencies to the Raspberry Pi x64 pre-built image already (av1-whip support excluded), but for other users you may need to ensure you have the gst-plugins-rs installed to get WHIP out working. This may also mean you’ll need Gstreamer 1.22 installed. If you want to use AV1, you’ll also need to ensure you have an AV1 encoder available within Gstreamer; there’s a few good options there.

The WHIP output support within Raspberry Ninja is added by means of the Gstreamer’s whipsink whepsrc Rust-based plugins.

To use, you can just do:

python3 --whip "" --test

The gst-launch-1.0 command line that’s printed to screen can be run on its own after, without Raspberry Ninja in cases, as Raspberry Ninja doesn’t need to make use of Websockets or other logic when dealing with WHIP.

You can test this out live using VDO.Ninja still, as VDO.Ninja supports WHIP playback without needing your own SFU or server. Just open in your browser FIRST, and then run the following command line:

python3 --whip "" --test

You should see your video play on the VDO.Ninja website within a few seconds after you start publishing, and there should be audio included.

Please note that you don’t need Raspberry Ninja to use WHIP, but Raspberry Ninja will handle all the Gstreamer pipelining, audio/video device detection, and I am open to feature requests.

As noted above, you can just re-use the gstreamer pipelines outputted by Raspberry Ninja when using WHIP output, but below I offer you a few test pipelines you can try from the command line to get you started there without Raspberry Ninja:

# To view the stream, FIRSTLY, open this link:

# to publish h264 to that viewer, use the following
gst-launch-1.0 videotestsrc ! videoconvert ! x264enc ! rtph264pay ! whipsink whip-endpoint=""

# or to publish av1 to that viewer, use the following:
gst-launch-1.0 videotestsrc ! av1enc usage-profile=realtime ! av1parse ! rtpav1pay ! whipsink whip-endpoint=""

In this case, we’re publishing direct to the browser, so it needs to be opened first, as we’re using VDO.Ninja to do the WHIP playback without WHEP.

You can find many more WHIP/WHEP options within VDO.Ninja, and a great place to start playing is at This page offers a web-based WHIP/WHEP player/publisher, as well as support for SVC, insertable streams, and all the advantages of VDO.Ninja in a simple to use web interface.

As for Meshcast support, Meshcast is getting WHIP-ingest support; the availability of this update is just pending some more testing and deployment.

Custom Gstreamer audio/video source pipeline

You can set your own custom video and audio device, as shown below, if you want more control over the inputs. The encoder portion is still handled by the code.

python --rpi --video-pipeline "v4l2src device=/dev/video0 ! video/x-raw,framerate=30/1,format=UYVY" --noaudio --debug

Avoid apostrophes in your pipelines, as that can cause a syntax issue, and you shouldn’t need to escape the brackets.

–video-pipeline and –audio-pipeline are available.

--format can specify what the encoder should be expecting, if that’s needed. ie: --format YUYV

NDI support

I’ve been experimenting with NDI support, but it’s not working properly yet; debugging needed still. Some users have it working though, although they haven’t contributed those changes. It will be a bit before I can get around to doing it myself. Probably not too hard if you wanted to give it a go and contribute to the project.

OpenCV / Tensorflow / FFMPEG / FDSink / Framebuffer support

There’s support for OpenCV/Framebuffer (–framebuffer STREAMIDHERE) and FDSink now. There’s a Youtube video link below demoing how to use Raspberry.Ninja to bring raw BGR video frames into Numpy.

This should allow you to use Tensorflow, OpenCV, or whatever Python script to access the webRTC video stream as raw frames. You can also use it to turn a webRTC stream into a motion-JPEG stream, useful if needing to publish to a device that doesn’t support webRTC. ie: Octoprint 3d printer server supports a MJPEG video preview link, but not a webRTC link.

Hardware options

Of the Raspberry Pi devices, the Raspberry Pi 4 or the Raspberry Pi Zero 2 are so far the best options on this front, depending on your needs. Any of the Nvidia Jetson devices should work fine, but only the Jetson Nano 2GB, 4GB, and NX have been tested and validated. If you wish to use other Jetson devices, you’ll need to setup and install Gstreamer 1.19 yourself on those systems, as no pre-built image will be provided at this time. (Unless someone wishes to donate the hardware that is) Any other Linux system or SBC embedded system is on the user to setup and install at this point, but they should closely follow the same steps that the Nvidia Jetson uses.

It’s rather hard to install everything needed on a Raspberry Pi Zero 2 directly, due to the limited memory, so I do recommend that if installing from scratch that you use a Raspberry Pi 4 with 4GB or greater.

Camera options

There’s plenty of options for the Rasbperry Pi and Nvidia Jetson when it comes to cameras and HDMI adapters. The easiest option for a Raspberry Pi is to use one of the official Raspberry Pi camera. These are normally just plug an play on both platforms and well supported.

USB cameras are options, but currently with Raspberry Pi devices these are only supported up to around 720p30. USB 3.0 devices are even less supported, as you need to ensure the Raspberry Pi you are using supports USB 3.0; for example, a Camlink will not work on a Raspberry Pi 3.

If low-light is important to you, the Sony IMX327 and IMX462 series of sensors might appeal to you. They are generally designed for security camera applications, but with the use of an IR Filter, you can make them adequate for use a standard video cameras. These options may require additional gstreamer and driver work to have work however, so they are for more advanced-users at this time.

I have gotten the low-light Arducam IMX462 to work with the newest image for RPI working (with the bullseye images). It might require a small change to the dtoverlay line in the /boot/config.txt file though to configure your specific camera, but I think I have most working now without any need drivers. (a few exceptions) , oh, and if you are changing dtoverlay, you might need to also comment out the camera auto detect link that is also in the config.txt file. (else it might not work)

Links for such low-light cameras: (I own this camera and it works on a Raspberry Pi 4 with my newest created RPi image. It works if you do not use the pivariety daughterboard and just connecting directly; you’ll need to change the config.txt file a bit and use –libcamera to use though) (USB-based; more compatible with other devices)

You can buy IR Filters, or you can buy lenses that come with IR filters, if needed, for pretty cheap. Many are designed for security applications, so be aware.

360-degree cameras

Support for the Theta 4k 360 USB camera has been added. Has been tested with the Jetson. It is likely too slow to use with a Raspberry Pi though.

Install script and brief usage example found here:

HDMI Input options

As per HDMI adapters, a 1080p30 USB 2.0 HDMI to MJPEG adapter can usually be had for $10 to $20, although there are many fake offerings out there. I’ve tested a $12 MACROSILICON HDMI to USB adapter, and it works pretty well with the Jetson (and OK with the RPI), although finding a legitimate one might be tricky. On a Raspberry Pi 4, 1080p30 is posssible with the HDMI to USB adapter, but audio currently then goes out of sync; at 720p though, audio stays in sync with the video more frequently. Audio sync issues might be resolved in the future with more system tuning.

There’s another option though, and that is to use an HDMI to CSI adapter for Raspberry Pis, such as the C780A ($29 USD), although the frame rate of an HDMI to CSI option is limited to 1080p25 (due to 2 CSI lanes only). It’s also slightly more expensive than the HDMI to USB alternative. The RPi Compute Module boards seem to have four-lanes of CSI available though, so 30-fps might be achivable there if you buy the compatible board (C780B ?)

Audio is also more challenging when dealing with the HDMI to CSI adapters, as you need to connect audio from the board via I2S to the RPi. This isn’t easy to do with some of the HDMI to CSI boards, but there are a couple options where this is a trival step.

Please note before buying that there are different HDMI to CSI2 boards, and they might look similar, but they are definitely not equal.

HDMI to CSI boards are not plug-and-play currently, as they do require a couple tweaks to the boot file at the very least, and maybe an update to the EDID file. (script provided for that). Depending on the video input signal, you might need to further tweak settings, such as colorimetery settings. This not really an issue with the HDMI to USB adapters, as they convert to a very standard MJPEG format, making them more plug and play friendly.

Please share with the community what works well for you and what did not.

MIDI options

When using the --midi parameter, video and audio are disabled. Instead, the script can send and recieve MIDI commands over VDO.Ninja. Supports plug-and-play, although you may need to install python-rtmidi using pip3 first.

Incoming MIDI messages will be forwarded to the first MIDI device connected to the Pi. Adding &midiout to the viewer’s view-link will have that remote browser send any MIDI messages (such as from a USB DJ Controller) to the raspberry_ninja script, which will then be forwarded to the first local MIDI device

Outgoing MIDI messages will be sent to connected viewers, and if those connected viewers have &midiin added to their view-links, those MIDI commands will be forwarded to the connected MIDI devices.

If using a virtual MIDI device on the remote viewer’s computer, such as loopMIDI, you can target that as both a source and target for MIDI commands. This is especially useful for connecting VDO.Ninja to DJ software applilcations, like Mixxx or Serato DJ Pro, which supports mapping of MIDI inputs/outputs.

Please note, the raspberry_ninja script can both send and recieve MIDI commands over a single peer connection, which is a bit different than how video/audio work currently. It’s also different than how browser to browser currently is setup, where a sender won’t ever request MIDI data, yet the raspberry_ninja code does allow the sender to both send and receive MIDI data.

midi demo video:



Discord Support

Support is available on Discord at in channel #raspberry-ninja