Frank

Frank is an adults-only companion chat. He is warm, direct, and sensual when you invite that, and the hard line does not move: consenting adults only. The conversation stays in the browser on the device. There is no account.

Boundaries

Frank refuses, in one sentence, and redirects when a request involves any of these:

• minors or age-play
• non-consent or coercion, including a stranger groping someone in public
• exploitation, violence, or illegal harm
• real, identifiable people

A server-side check rejects obvious cases before the model is called. It is a light, high-precision filter so ordinary adult talk still goes through. It is not a substitute for the system prompt, which keeps the same boundaries in every tone.

Tones (sweet, playful, spicy) only change delivery. They do not loosen the boundaries.

Setup

Node 22.

npm ci
npm run dev

The dev server listens on port 8080. The first screen is an 18+ confirmation, remembered on the device. After that, the chat is the whole app.

Environment

Set these on the server. Do not prefix them with VITE_ and do not commit them.

Variable	Required	Default	Purpose
XAI_API_KEY	To reply		Server-only xAI API key. On a Grok deploy the platform injects it.
XAI_MODEL	No	grok-4.5	Chat model id (letters, numbers, ., _, -). Anything else falls back to the default.
XAI_VISION_MODEL	No	grok-4.7	Model for a turn that includes a photo. See below.
XAI_TEMPERATURE	No	0.7	Sampling temperature, clamped to 0–2. Invalid values fall back to the default.

If XAI_API_KEY is missing, Frank stays on screen and says the model is unavailable. Messages already on the device can still be read. Nothing is sent to xAI.

Replies stream from POST /api/frank. Output is capped (max_tokens 450) so a visitor cannot run an unbounded completion.

Photos

A message can include up to four photos: jpg, png, webp, or gif, from the photo picker, a drop, or a paste. Large stills are resized on the device (long edge 1600px, jpeg) before they are kept. Each photo has to stay under about 1.4 million characters of data URL. A gif under 900KB stays animated in the thread. Anything over the cap shows an error and is not sent.

Frank sees the photo. The request uses xAI chat-completions image parts (image_url with a base64 data URL, detail: auto). xAI documents jpeg and png only, so a webp or gif is turned into a jpeg still on the device before that call. The animated gif in the thread is unchanged.

Image input was checked against docs.x.ai on 2026-10-29. These chat model ids accept images:

• grok-4.7 (text and image in, text out)
• grok-4.6, grok-4.5, grok-4.5-latest, grok-4.3, grok-4.3-latest (text, image → text)
• grok-4 (chat-completions image examples)
• grok-2-vision-1212 and other ids that contain vision (legacy)

grok-4.7 does accept images. A photo turn uses XAI_MODEL when that id is one of the above. Set XAI_VISION_MODEL to force a different model for photo turns. If that variable is unset and XAI_MODEL cannot take images, photo turns use grok-4.7.

Only the four newest photos in the thread are sent again on a later turn, so a follow-up can still refer to a recent picture without uploading the whole history. Older photos stay visible on the device. Tapping one opens it full screen: Esc, the close button, or a tap on the dark area dismisses it; arrows and a swipe move between photos in the thread; pinch, double-tap, or the scroll wheel zooms.

Photos are not written on the server and are not logged. They live in the same on-device thread as the text. Clearing the chat removes them. A very long thread drops the oldest photos from device storage before the text.

Clips

A message can include one short video: mp4, webm, or mov, up to 10 seconds and 20 MB. The clip is not uploaded and is not kept. The browser takes up to four stills and sends those through the same photo path, so Frank sees the clip as frames. The stills stay in the on-device thread with the photos.

Chat models on xAI accept images, not a video file, in the reply request. The stills are the way a clip reaches Frank.

Voice

The microphone turns a voice note into text in the composer. With XAI_API_KEY, that note is transcribed by xAI speech-to-text (POST /v1/stt) and you can edit it before sending. Without the key, browsers that support the Web Speech API dictate straight into the composer. Notes are capped at about 20 seconds. The audio is not stored.

The speaker control speaks Frank's replies. With the key, speech uses xAI text-to-speech (POST /v1/tts) in the voice chosen at the top of the chat (Rex by default). If that call fails, the browser's own speech synthesis speaks the line. A reply that crosses the hard boundaries is not spoken. "Hear it" under the latest reply plays it again. The voice and the speaker switch are remembered on this device.

Scripts

npm test
npm run lint
npm run typecheck
npm run build

npm test runs the script tests and Frank's boundary, config, stream, and storage tests.

Deploy

The production build is Vite plus Nitro's Vercel preset (npm run build). vercel.json sets the install command. Configure XAI_API_KEY, and optionally XAI_MODEL, XAI_VISION_MODEL, and XAI_TEMPERATURE, in the Vercel project for Production and Preview.

Do not commit .vercel/output, preview logs, screenshots, or secrets. .env files are gitignored. The platform injects XAI_API_KEY for a Grok-hosted deploy, so a local .env is not part of this repo.

Privacy

The 18+ confirmation, the tone, and the message history — including photos — live in localStorage on the device. Clearing the chat removes the thread from that store. The confirmation stays until the site data is cleared. The server does not store photos; it only passes them through on the reply request.