EagleFeeder (MP4)

Video changed ComfyUI's output story. These days the real video models - Wan, LTX, the whole crop - emit a native VIDEO tensor, not just a stack of frames, and newer ComfyUI ships a proper SaveVideo node for it. EagleFeeder (MP4) is that node with a different ending: instead of dropping an MP4 in your output folder, it encodes the video and feeds the file straight into a running Eagle app, tagged and filed into a folder you name. It's the third node in the ComfyUI Eagle Feeder pack, and the one with the most subtle trap: it takes a VIDEO input, not an IMAGE batch.

If you're coming from the other two nodes in the pack (PNG, animated WebP), everything about the Eagle side is identical - same local file server, same API calls, same folder lookup. The difference is all in the video plumbing, and it's where people get stuck.

How it works

The mechanism is the pack's one trick, repeated: the first time any Eagle Feeder node loads, it starts a tiny local file server on port 8000 serving a scratch folder inside the pack. On run, the node asks Eagle's /api/folder/list for your folder's ID, saves the video via ComfyUI's native VideoInput.save_to(), then calls Eagle's /api/item/addFromURL with a http://localhost:8000/<filename> URL, your comma-separated tags, the folder ID, and your token. Eagle fetches and files it. Scratch files older than two days are purged automatically.

Note the save_to() call - that's ComfyUI's bundled video encoder doing the work, so this node inherits whatever video support your ComfyUI build ships (ffmpeg and friends). It also means you need a recent ComfyUI. The node imports comfy_api.input.VideoInput and comfy_api.util.video_types under the hood; on an older install those don't exist and the node won't even register. If it's missing from your node list after install, update ComfyUI first, then blame the pack.

The inputs that matter

• video - the VIDEO tensor from a video-model pipeline. This is the trap: unlike SaveVideo's frame-batch ancestors, you don't feed it a list of images. Wire it from whatever produces native video output in your graph.
• format and codec - both default to auto, with mp4 / h264 as the explicit alternatives. Auto does the right thing in practice; set them explicitly only if you're chasing a specific container or codec.
• tags, folder_name, eagle_host, eagle_token - the shared Eagle plumbing. tags is comma-separated and forceInput, so wire your prompt or a tagger output straight in. folder_name must already exist in Eagle - no auto-create, empty means no folder, duplicate names pick arbitrarily. eagle_host defaults to http://localhost:41595 (Eagle's fixed API port) and the token is under Preferences → Developer.
• embed_workflow - on by default, and for MP4 it attaches the workflow as video metadata. Handy given the pack's whole point is curating a library you can actually use later.

Install

ComfyUI Manager, search "ComfyUI Eagle Feeder," or:

cd ComfyUI/custom_nodes
git clone https://github.com/tighug/comfyui-eagle-feeder.git

Restart. No model downloads. Mild gotcha: the pack's requirements.txt pins torch==2.6.0 exactly, which can clobber a ComfyUI install already on a different torch build - skip the dependency install if Manager wants to swap torch; the nodes run on whatever ComfyUI already has.

Where it trips people

Same three as the rest of the pack: Eagle must be running or the folder lookup dies; a missing token kills every upload; and port 8000 has no fallback, so if it's occupied the file server dies silently and you get connection-refused uploads. Then there's the video-specific one - feeding it an IMAGE batch instead of a VIDEO tensor, which fails or silently misbehaves. And if Eagle lives on another machine, remember the URL it gets is always localhost:8000, so you need an SSH tunnel (or shared container network) making that port reach your ComfyUI box. Get past those and it's a genuinely tidy end-of-pipeline node for the video workflow.