You can run computer vision on an OBS broadcast without a plugin, a GPU or any new hardware. Point OBS’s Custom stream setting at the server URL and stream key from Roboflow Cloud Streaming, then describe what you want tracked to the Roboflow Agent, and it builds and deploys a Workflow that reports those stats live.
OBS (Open Broadcaster Software) is where a lot of live video and media broadcasts are mixed, composed, and put together in a single broadcast. Used by streamers, newsrooms, schools, sports leagues, and hobbyists, OBS is one of the most popular streaming/broadcasting tools in the world.
Streams, especially livestreams, however, are hard to keep an eye on. Knowing what is on screen at any moment means having someone watch the whole thing as it airs, which gets harder the longer the stream runs and the more streams you have. Vision AI can help reduce that workload while making things more deterministic.
Today, we’ll test out running real-time and continuous inference on a live stream with no additional hardware or software, using Roboflow Cloud Streaming.
Run Vision Models on OBS Live Broadcasts with Roboflow
You can run vision models on an OBS broadcast in three steps: create an RTMP source in Roboflow Cloud Streaming, point OBS at it, and deploy a Workflow to the stream.
Step 1: Create a Cloud Streaming Source
Open Roboflow Cloud Streaming under the Deployments pane, then click on “Add Video Source”. Alternatively, you can tell the Roboflow Agent or your own AI Agent (Claude, Codex, ChatGPT, etc) that “I want to process an RTMPS stream on Roboflow Cloud Streaming”
Regardless of how you choose to connect, you should be prompted to select a method and name your source. OBS can stream over RTMP or its encrypted version, RTMPS, so we’ll select the “RTMP Stream” option.
Once created, you’ll be provided with a Server URL and a Stream Key. Hold on to these for the next step. Treat the Stream Key like a password and don’t share it with others.
Step 2: Configure a Stream
Now that we have the stream details, go to OBS and, within OBS’s preferences page, navigate to the stream menu and, under the service dropdown, click “Custom”.
Input the Server URL in Server and copy your Stream Key in.
Then start the stream by clicking “Start Streaming” within your OBS Controls.
That’s it. With two steps, you can now stream your broadcast into Roboflow to run computer vision models and workflows on your OBS livestream. Give it a few seconds, and your broadcast should show up on the source in Cloud Streaming.
Step 3: Run a Workflow
If you’re using Roboflow Agent, it’ll then prompt you with what you want to do with the stream. Describe it in plain words, the way you would explain it to a coworker. Example Workflow
On this VOA broadcast that we piped in through OBS, let’s say we wanted to do some broadcast analytics on three different statistics over the past 5 minutes:
The average number of people shown within each frame
The median screen space (percentage of the total frame) one person occupies
For fun, we’ll also keep a stat of the % of people who have suits on. After telling the Agent to work on that for a bit and answering some questions, we now have a live stream output. A pretrained model finds the people, and SAM 3 finds the suits from a text prompt, so there was nothing to label or train.
Once the Agent deploys the workflow to the stream we set up earlier, we can get a live feed:
Start With Roboflow Cloud Streaming
In this guide, we connected an OBS broadcast to Roboflow Cloud Streaming and then ran a workflow that reports live statistics about who is on the screen in real time. The setup works for any other intelligence you want to extract from a broadcast. Because the source is a standard RTMP stream, any encoder that can push RTMP or any of the other protocols we support will work with Cloud Streaming.