An open-source, dual-mode AI chat component you can drop into any site and point at your own story. It renders as a collapsed hero input that expands into a full conversation overlay — one React component, two modes, no separate widget. EU-region by default, runs on hobby-tier infrastructure, MIT licensed.

Live demo: https://applicaiton.vercel.app

Built on the foundation of Santiago Fernández's cv-santiago — the dual-mode chat idea and the thin-embed pattern started there. Forked forward, themed, and extended.

• Dual-mode chat — one component: a collapsed hero input, and the full conversation overlay it expands into. The component owns its own embedding (scroll lock, focus, close); host pages only control layout.
• A theme system — visual identity lives entirely in CSS custom properties. Change one data-theme attribute and the whole page recolors, the chat included (see Theming).
• A per-company context loader — private calibration files in Upstash, matched by keyword and lifecycle-gated. Your targeting never lives in git.
• A prompt-stuffed knowledge base — markdown files, cached, no RAG and no vector database. The whole story fits in the context window.
• A thin embed — the host page is three lines.

A single request, server-side: rate limit (Upstash sliding window per IP) → assemble the system prompt from the knowledge base → match a per-company context file if a keyword fires → stream the response from Claude Haiku 4.5 via the Vercel AI SDK. Every turn is traced in Langfuse (EU region). No vector database, no retrieval step.

git clone https://github.com/derheissenberg/applicaiton
cd applicaiton
npm install
cp .env.local.example .env.local   # then fill in your keys
npm run dev

Open http://localhost:3000. Edit app/page.tsx to change the demo page.

Set these in .env.local — see .env.local.example for the full list:

• ANTHROPIC_API_KEY — required. As shipped the bot streams from Claude Haiku 4.5. Because it goes through the Vercel AI SDK (provider-agnostic), pointing it at another provider — OpenAI, Google, and so on — is a small swap in app/api/chat/route.ts, not just an env change.
• UPSTASH_REDIS_REST_URL and UPSTASH_REDIS_REST_TOKEN — optional (rate limiting). Unset, the limiter fails open and every request is allowed.
• Langfuse keys — optional (tracing). Unset, tracing is skipped.

The Anthropic key is the only thing you need to run locally.

Do not wrap values in quotes. Next.js treats the quotes as part of the value, and several of these dependencies fail open silently when that happens.

Forking ApplicAIton to point at your own story is a handful of edits — no component code required. Everything personal lives in two files you own: the knowledge folder and the persona prompt.

1. Swap the knowledge base. Replace the markdown in knowledge/ with your own — identity, narrative, experience, projects, skills, credentials, case studies, philosophy, FAQ, guardrails. This is the bot's only source of truth; it answers from here and nowhere else. See Knowledge base.
2. Retarget the persona prompt. chatbot-prompt.txt names its subject throughout — but you don't need to rewrite it. Retarget only the sections that mention a person: §1 Identity, §2 Source of truth (name + contact email), §6 positioning, and §12–13 (contact email + portfolio links). The voice, formatting, anti-injection, and honesty rules (§3–5, §8–11) carry over unchanged.
3. Pick or define a theme. Use a shipped identity (dark-tokyo or stefan-portfolio) via the theme prop, or add your own [data-theme="your-name"] palette. No color is hardcoded in component code. See Theming.
4. Set one key. Only ANTHROPIC_API_KEY is required to run. See Environment.
5. Embed it. Three lines on your host page. See Embedding.
6. Deploy. Vercel or any Next.js host. See Deploy.

Per-company calibration files are optional and never live in git — see Per-company context.

The host page controls only layout; the component brings everything else.

import { Chat } from "@/components/chat/Chat";

<section style={{ ["--chat-hero-min-height-desktop"]: "100dvh" }}>
  <Chat theme="dark-tokyo" />
</section>

Visual identity is CSS custom properties — no hardcoded colors, and no theme branching in component code.

• Hero height is set by the host page, not by themes. Pass --chat-hero-min-height-desktop on a parent element via inline style or CSS. The default is auto (content-fit). Mobile hero height stays theme-controlled via --chat-hero-min-height-mobile (default 100dvh).
• The chat ships two identities — dark-tokyo (the demo default) and stefan-portfolio. Select one with the theme prop.
• Page-wide theming — a shared color-token contract lives on a data-theme wrapper around the page. Change that one attribute and the whole surface recolors: the chat hero and everything around it move together. The chat's color tokens reference the shared variables with built-in fallbacks, so a bare <Chat /> with no wrapper still renders correctly on its own defaults. Define your own palette by adding a [data-theme="your-name"] block of variables.

The bot's knowledge is plain markdown in /knowledge (~25K tokens, prompt-stuffed with caching). Replace those files with your own and the bot answers from your story. The component code never hardcodes personal facts — you swap the folder, not the source. The one thing the folder doesn't carry is the bot's name, contact email, and positioning: those live in the persona prompt (chatbot-prompt.txt), so retarget it too — see Make it yours.

Beyond the base knowledge, you can write per-target calibration files stored in Upstash under ctx:* keys. On cold start the loader scans them, parses their frontmatter, drops anything not marked active, and substring-matches incoming messages against their keywords; the first match is appended to the system prompt for that conversation and the behavior shift is silent. Targets are never discoverable from the repo — the trigger keyword is the only boundary. The authoring method is documented as a skill so you can write your own without leaking who you're writing them for.

A suite of 21 cases across six areas — boundary testing, company context, factual accuracy, persona adherence, response quality, and safety/jailbreak — runs as a CI gate on every push. The bot speaks about its subject in the third person and refuses salary specifics, opinions on former employers, personal-life questions, and prompt-leak attempts.

Runs on Vercel Hobby, the Upstash free tier, Langfuse Hobby, and a Haiku-class model — roughly €0–8/month at demo volume.

Built on the foundation of Santiago Fernández's cv-santiago. The dual-mode chat idea and the thin-embed pattern started there; the lineage is the point, not a footnote.

MIT. Fork freely, make it yours.

The easiest way to deploy is the Vercel Platform. See the Next.js deployment docs for details.