atlas-explainer

Turn one topic into a finished, narrated, subtitled, scored papercraft explainer video (Vox / Kurzgesagt style) — end to end, from Atlas Cloud APIs + local ffmpeg. No editor, no timeline, no GPU.

This is a self-contained agent skill (works in Claude Code and any skill-aware agent) plus a plain CLI. Point it at a topic; it storyboards, generates the footage, narration and music, bakes on titles and subtitles, and assembles an mp4.

pip install imageio-ffmpeg pillow
export ATLASCLOUD_API_KEY=apikey-xxxx      # https://www.atlascloud.ai/console/api-keys

python3 scripts/run.py /tmp/myvideo --topic "how the shipping container reshaped global trade"
# → /tmp/myvideo/out/final.mp4   (~$0.95, ~4 min, 6 scenes, 1280x720)

Options: --scenes 6 · --vertical (9:16) · --lang en · --force.

Install as a Claude Code skill

git clone https://github.com/AtlasCloudAI/atlas-explainer ~/.claude/skills/atlas-explainer

Then just ask your agent: "make a 1-minute explainer about X" — it reads SKILL.md and drives the whole flow. Full walkthrough (Chinese): HOWTO.md.

What's in the pipeline

Step	Model / tool (all via Atlas Cloud)
Storyboard	any Atlas LLM (plan.py) — or you/your agent hand-author scenes.json
Scene footage	google/gemini-omni-flash/text-to-video
Narration	xai/tts-v1 (or bytedance/seed-audio-1.0 for Mandarin)
Music	minimax/music-2.6
Titles / subtitles / assembly	local PIL + ffmpeg (bundled via imageio-ffmpeg)

Self-contained & portable

Only external deps are pillow and imageio-ffmpeg — everything else is Python stdlib.

• ffmpeg comes from imageio-ffmpeg (no system install); fonts resolve across macOS/Linux/Windows with a graceful fallback.
• Networking tolerates a broken/absent proxy (env-proxy → direct fallback) and retries transient drops.
• A generated scene's output URL is saved before download, so a failed download re-downloads instead of re-billing the model.

Notes before publishing a film

• No real, recognizable people. gemini-omni-flash moderation blocks real-person likenesses. Papercraft / objects / abstract only. (For real-person collage, that's a different tool.)
• No text baked into the AI footage — all words are overlaid afterward (AI-rendered text garbles).
• Never commit your key; scripts/atlas_api.py reads ATLASCLOUD_API_KEY from the env.

Atlas MCP note

The Atlas MCP has no audio-generation tool yet, so this skill calls generateAudio directly. A ready-to-merge atlas_generate_audio MCP tool is in MCP_AUDIO_TOOL.md.

---

MIT. Built to drive Atlas Cloud API usage — fork it, ship films. Sibling project (image-collage motion pipeline): atlas-collage-studio.