Apppublic

BKarthikeyan/browser-agent-stream

sourceHugging Facemitupdated 3mo agoView on Hugging Face

0likes

App README

Zed — Computer-Use Agent

A live computer-use agent with full desktop streaming via WebRTC (Selkies-GStreamer).

What runs inside

Component	Tech	Why safe on HF
Virtual display	Xvfb	Standard X.Org, never flagged
Window manager	openbox	Lightweight, no remote-desktop protocol
Streaming	Selkies-GStreamer (WebRTC H.264)	GStreamer only, zero VNC binaries
Browser control	Chromium CDP	No VNC, pure DevTools Protocol
Agent API	FastAPI + uvicorn	Standard Python web framework
Reverse proxy	nginx on :7860	HF-required single port

Endpoints

URL	Description
`/`	Screenshot polling viewer (instant fallback)
`/stream/`	Live WebRTC desktop stream
`/screenshot`	Single JPEG screenshot
`/health`	Health check
`/agent/inject`	`POST {"instruction": "..."}` — run agent task
`/agent/execute`	`POST {"action": "..."}` — single action
`/webrtc/signal`	HTTP-based WebRTC signaling proxy

HF Space Variables (Settings → Variables)

TURN_HOST   = staticauth.openrelay.metered.ca   # or your Render coturn
TURN_PORT   = 443
TURN_SECRET = openrelayproject                  # or your secret
LLM_API_KEY = <your API key>
LLM_BASE_URL = https://server-llm-1.onrender.com/v1
LLM_MODEL   = auto

Local development

bash

cd desktop-agent/huggingface-space
docker build -t zed-agent .
docker run -p 7860:7860 \
  -e TURN_HOST=staticauth.openrelay.metered.ca \
  -e TURN_PORT=443 \
  -e TURN_SECRET=openrelayproject \
  -e LLM_API_KEY=<key> \
  zed-agent

Open http://localhost:7860 → screenshot view Open http://localhost:7860/stream/ → live WebRTC desktop