NarrateBot 3000

Your world, narrated.

Live: https://narratebot3000.kelvinct.com · accessible view: https://narratebot3000.kelvinct.com/?view=clean

NarrateBot 3000 turns any webcam into a live narrator that describes the world for blind and low-vision people, in whatever language they speak, with nothing to install. It also does nature documentaries and comedy movie trailers, because why not.

Run locally

vercel env pull .env.local   # pulls the project's env vars (git-ignored)
vercel dev                   # serves index.html + /api on http://localhost:3000

The webcam needs localhost or HTTPS. To test on a phone, use the deployed link.

Access

The hackathon is over, so the server's own Gemini and ElevenLabs keys are reserved for the owner.

• Bring your own keys. Pressing ORDER NOW / Start without access opens a "Use your own keys" dialog: a Gemini API key (required, checked with Google before it's accepted) and an ElevenLabs API key (optional; without one, the browser voice speaks). The (i) buttons explain where to get each key. Keys are kept only in an encrypted, HttpOnly cookie in that browser (nb3k_session, AES-256-GCM, 30 days), sent with each request and never stored or logged on the server. Settings → Access shows the current access and has "Change keys" and "Remove my keys". Visitors' usage doesn't count against the daily caps below; per-device rate limits still apply.
• Owner. Tap the "AS SEEN ON TV!" starburst 5 times within 3 seconds (in the accessible view, tap the "NarrateBot 3000" heading), or press Shift+O, then enter OWNER_PASSCODE. Attempts are limited to 5 per 15 minutes per IP. Owner sessions use the server keys and the daily caps; "Lock owner access" in Settings signs out. Changing OWNER_PASSCODE signs out every owner session; changing SESSION_SECRET signs out everyone.

Environment variables (names only)

Name	Purpose
GEMINI_API_KEY	Gemini vision for owner sessions
ELEVENLABS_API_KEY	ElevenLabs voices for owner sessions (optional; the browser voice is the fallback)
OWNER_PASSCODE	Owner password (a long random string). Unset = no owner access
SESSION_SECRET	Encrypts the session cookie (at least 32 characters; 96 hex recommended). Unset = no sessions at all
VOICE_DOCUMENTARY, VOICE_TRAILER, VOICE_DESCRIPTIVE, VOICE_PUMPKIN	Override the default voice IDs in api/speak.js
VOICE_SPANISH, VOICE_CHINESE	Optional per-language voices
GEMINI_MODEL_FUNNY, GEMINI_MODEL_FAST, ELEVENLABS_MODEL	Optional model overrides
ELEVENLABS_MODEL_TELUGU	Optional; Telugu narration uses eleven_v3 by default because eleven_multilingual_v2 has no Telugu
DAILY_NARRATION_LIMIT (900), DAILY_TTS_CHAR_LIMIT (20000)	Daily caps on the owner's keys
KV_REST_API_URL + KV_REST_API_TOKEN (or UPSTASH_REDIS_REST_URL + UPSTASH_REDIS_REST_TOKEN)	Upstash Redis for shared rate limits

Voice settings (stability, style, and so on) live in the VOICES object at the top of api/speak.js. Prompts live in api/_prompts.js.

Narration languages: English, Spanish, Chinese, Hindi, Telugu, French, Arabic, Korean, Japanese, Portuguese, plus a custom one from Settings. Non-English lines also come back with an English translation, shown under the subtitle when the site is in English. ElevenLabs audio streams to the page (GET /api/speak, with the session cookie), so speech starts before the whole clip is generated. With a visitor's own ElevenLabs key, a voice their plan can't use is retried once with the premade "George" voice. Speech is requested sentence by sentence, using ElevenLabs request stitching (previous_text/next_text) so the sentences sound continuous, and the next line is captured and its first sentence loaded while the current one plays.

Hotkeys

Key	Action
1 2 3 4	Documentary, Funny, Descriptive, Pumpkin
L	Cycle narration language (any mode; also the 🌐 picker in the control bar)
N	Narrate now
S	Force stillness jokes on/off
T	Dark/light theme
A	Accessible view
Space	Pause/play
K	Settings

Tech

Gemini (vision), ElevenLabs (voices), vanilla JS/HTML/CSS with no build step, Vercel serverless functions, and Upstash Redis rate limiting.

Credits

Background music (documentary.mp3, funny.mp3) is from Pixabay, used under the Pixabay Content License. The tracks aren't covered by this repository's code and can't be redistributed on their own.

License

Copyright (C) 2026 Kelvin (@kelvintiger).

The code is licensed under the GNU Affero General Public License v3.0 or later. You can use, change and share it, but any copy or modified version, including one you only run as a website, must keep this copyright notice, credit the original project, and publish its full source code under the same license. The background music isn't covered by this license (see Credits).