Briefing video from 3 news sources — one shareable MP4.

An editor pastes three URLs. The agent reads them, writes a neutral 60-word brief, voices it in a premium anchor cadence, generates four scene-relevant B-roll clips, stitches them, and mixes the audio. Output: one social-ready video — not two broken artifacts.

Final deliverable

Single MP4. Click play — visuals follow the script, voiceover follows the visuals. The whole thing is one drag-and-drop into X / LinkedIn / Reels.

The four B-roll scenes inside

Each generated to match a specific beat in the narration. seedance-i2v-fast animates flux-dev stills; ffmpeg-concat stitches them; ffmpeg-mux drops the inworld-tts voice over the top.

What the agent read

Three real URLs; obscura's headless browser pulled the article body of each:

What the agent wrote

One gemini-text call synthesizes a neutral 60-word brief from the three article bodies:

"Breaking. NVIDIA partners with Corning on a major U.S. manufacturing push for optical connectivity — three new facilities, three thousand high-paying jobs. Meanwhile, Anthropic locks in three hundred megawatts and two hundred twenty thousand GPUs through a new compute deal with SpaceX. The agentic AI infrastructure war just escalated."

Voiced with inworld-tts — the production-grade anchor preset. Not the lightweight chatterbox default.

The full pipeline — 11 MCP / SDK calls

# 1-3. Fetch the three sources via obscura (parallel)
obscura-extract-markdown {url: marketingprofs.com/...}
obscura-extract-markdown {url: cnbc.com/...}
obscura-extract-markdown {url: devflokers.com/...}

# 4. Synthesize the neutral 60-word brief
gemini-text {prompt: "summarize these 3 articles neutrally..."}

# 5-8. Four scene-relevant stills (parallel) — newsroom + factory + data center + GPU
create_media action=generate model=flux-dev × 4

# 9-12. Animate each still (parallel) — seedance-i2v-fast, ~7s each
create_media action=animate model=seedance-i2v-fast × 4

# 13. Premium anchor voiceover
create_media action=tts model_override=inworld-tts

# 14. Stitch the 4 clips into one 26-second video
ffmpeg-concat {clips: [newsroom, factory, datacenter, gpu]}

# 15. Mix the voiceover onto the concatenated visuals — the production cut
ffmpeg-mux {video_url: concat.mp4, audio_url: tts.wav}

Why this is unique on the SDK

• Multi-source synthesis. No Canva / Adobe Express tool reads three URLs and writes a brief. They all start from copy you've already written.
• Visuals match the script. Four scene-relevant clips, one per beat — not a single static hero on loop.
• Single shareable file. ffmpeg-concat → ffmpeg-mux produces one MP4 that goes straight onto social. No "open this audio + play this video" hand-off.
• Voice that doesn't embarrass you. inworld-tts anchor preset vs the chatterbox robot — production grade.
• Cost. Three URLs in, broadcast-ready briefing out, ~$0.55. Daily news series at this price is feasible.

Cost breakdown