Imported from luxonis/oak-examples (streaming/mjpeg-streaming/AGENTS.md). Install upstream with npx skills add luxonis/oak-examples --skill mjpeg-streaming. Copyright stays with the author.

AGENTS.md

Summary

This is the repository reference for HTTP MJPEG streaming with inline YOLOv6 overlays. Use it when you need the simplest browser-consumable stream in this category.

Use This Example When

• You need a plain HTTP MJPEG stream.
• You want detections drawn directly into the streamed frames.
• You need camera or replay input without a separate frontend app.

Do Not Use This Example When

• You need RTSP, WebRTC, or raw TCP instead of HTTP MJPEG.
• You need a standard DepthAI Visualizer topic workflow.
• You need a configurable server port or a control backchannel.

Quick Facts

• Category: streaming/mjpeg-streaming
• Shape: script+standalone
• Primary task: serve detected frames as an MJPEG stream over HTTP
• Entrypoint: main.py
• Standalone path: oakapp.toml
• Frontend: none
• Runs on: RVC2 peripheral, RVC4 peripheral, and RVC4 standalone packaging
• Requires: camera input or replay media, HTTP access to the runtime, and YOLOv6 model assets
• Input: live camera by default or ReplayVideo via --media_path
• Output: HTTP MJPEG stream on http://<host-or-device>:8083
• Models: YOLOv6 YAMLs in depthai_models/
• Visualizer / UI: browser or any MJPEG-capable HTTP client

Read First

• README.md
• main.py
• utils/mjpeg_streamer.py
• utils/server.py
• utils/arguments.py

Architecture

• main.py builds a camera-or-replay input path and runs YOLOv6 through ParsingNeuralNetwork.
• utils/mjpeg_streamer.py is a HostNode that draws detections into the passthrough frame.
• That same host node starts a threaded HTTP server and continuously updates the frame served to clients.
• utils/server.py emits a multipart image/jpeg response for any GET request.

Constraints

• The HTTP server port is fixed at 8083 in utils/mjpeg_streamer.py.
• There is no dai.RemoteConnection or separate topic output; the stream itself is the main interface.
• In standalone mode the MJPEG server runs on the device-side app container, so clients must connect to the device IP rather than localhost.

Related Examples

• streaming/rtsp-streaming: use this when you need a standard RTSP endpoint instead of MJPEG over HTTP
• streaming/webrtc-streaming: use this when you need a browser-oriented interactive stream with runtime options
• streaming/poe-tcp-streaming: use this when you need a custom TCP transport and a control backchannel

Validation

• Run: python3 main.py
• Replay mode: python3 main.py --media_path <VIDEO>
• Success looks like: opening http://localhost:8083 in peripheral mode shows a live MJPEG stream with YOLO detections drawn on top
• Common failure meaning: the HTTP port is unreachable, replay media is invalid, or the user expects a Visualizer topic instead of a standalone MJPEG endpoint