Antzoki

Point Claude Code at a web app. Get a narrated 4K demo video.

No screen recorder. No editor. No re-shoots. One command turns a running web app into a polished, narrated product demo, and a small JSON file keeps it in sync forever.

/antzoki http://localhost:3000 my-product

Claude Code studies your app, asks you a few sharp questions, writes the script and the storyboard, then records, narrates, and composites a finished video while you do something else.

The name is Basque for "theater": the place where a show is staged.

---

The pipeline

demo.json ──▶  tts  ──▶  record  ──▶  build  ──▶  my-product-demo.mp4
 scenes        voice     Playwright   ffmpeg        3840 x 2160
 + UI steps   (ElevenLabs) page-only   one master    narrated
              raw + cached  4K capture  + ducked bed   broadcast loudness

Four stages, two data files, zero manual editing. The voice is generated first, the recording is paced to the voice, and the whole program is mastered once at the end. That last discipline is what kills the loudness jumps that make automated demos sound cheap.

Why it is different

Screen recorders capture you fumbling through a UI. Slideshow tools fake the product. Antzoki does neither.

• Code-aware. A subagent reads your routes, components, and test ids to script real selectors and real features, not guesses.
• A voice that is neither robotic nor salesy. ElevenLabs multilingual_v2 with request stitching and a neutral, documentary voice, mastered to one broadcast loudness.
• It interviews you. It asks about audience, story, and branding before writing a single line.
• Re-renderable, not re-recorded. Your demo is a demo.json. When the UI changes, you edit the affected steps and re-run. Cached voice clips mean only the scenes you changed re-synthesize. Commercial screen recorders cannot do this.

Install

The repo is a single-plugin marketplace:

/plugin marketplace add ark-232/antzoki
/plugin install antzoki@antzoki

Then, from inside any project:

/antzoki http://localhost:3000 my-product

If the target is omitted, it will ask. It confirms the app is reachable before doing anything.

Requirements

• Node 20+, ffmpeg, and poppler (pdftocairo, only if your demo shows PDFs).
• Playwright Chromium: cd pipeline && npm install && npx playwright install chromium.
• An ELEVENLABS_API_KEY for the production voice (cp .env.example .env). Without it, the macOS say voice is used, which is perfect for iterating.

Run the pipeline directly (no plugin)

Everything is also plain scripts against a demo.json:

cd pipeline
npm install && npx playwright install chromium

# fast iteration with the local voice, no key, no credits
TTS_PROVIDER=say node record.mjs ../examples/example/demo.json

# production voice + composite (key loaded from .env, never printed)
node --env-file=../.env record.mjs ../examples/example/demo.json
node --env-file=../.env build.mjs  ../examples/example/demo.json
# -> ../examples/example/out/example-demo.mp4

SKIP_VIDEO=1 node build.mjs ... rebuilds the audio without re-encoding the video.

The demo spec

A demo.json has an app block, a brand block for the title cards, and an ordered list of scenes. Each scene has a narration text and a list of declarative steps (click, fill, softNav, generate, showAsset, and more). Cards have no steps.

• schemas/demo.schema.json for the full shape
• skills/antzoki/reference/storyboard-actions.md for the action vocabulary
• examples/example/demo.json for a complete, working template

Pick a voice

cd pipeline
node voice-audition.mjs search "documentary narration"
node voice-audition.mjs sample <voiceId>:Name <voiceId>:Name

Play the samples, set voice.voice in demo.json. Prefer documentary or news voices, keep style at 0.

Tests

cd pipeline && node --test

Unit tests cover variable interpolation, locator resolution, the timeline math, the demo validator (it validates the bundled example), and the TTS config and cache key. CI runs them on Node 20, 22, and 24.

How it works

A full technical write-up is in docs/ARCHITECTURE.md: recording, voice, video and audio compositing, and the gotchas that took real debugging (single-master loudness, the alimiter auto-level trap, the xfade timebase fix).

Prior art

Inspired by the Claude Code ecosystem and projects like splitbrain/ndemo and digitalsamba/claude-code-video-toolkit. Antzoki's distinctive bets are the code-aware storyboard, the neutral mastered voice, and the re-renderable spec.

License

MIT.

_{Made by ark-232.}