Customer support · checked 2026-10-19

Can I vibecode MagicChat?

YES· one-shottable

A chatbot that answers from your own pages is a small retrieval pipeline: crawl, chunk, embed, fetch the closest passages and let a model answer from them, plus a widget on the page. An agent builds that in one sitting on SQLite and any OpenAI-compatible endpoint. What you give up is the hosted upkeep: scheduled re-crawls, abuse limits on a public endpoint and the ready-made embeds for Shopify or WordPress.

$29/moprice

$348you'd save a year

one sittingbuild time

—people replaced it

Before you start

What you need on hand.

DifficultyIntermediate · Some system setup or keys.

TimeOne to two hours for the first version

Claude Code, Codex or CursorAny of them can run the prompt.
Node.js 22better-sqlite3 and sqlite-vec are native modules, so build tools must be installed.
An LLM and embedding endpointAn OpenAI-compatible key, or a local server such as Ollama.
A public addressThe widget has to reach your server over HTTPS: a small VPS or a tunnel.

Before you buildMagicChat Free covers one bot with 50 pages and 100 messages a month. For a small site with little traffic that may be enough.

1. Install one agent: Claude Code, Codex, Cursor.
2. Make an empty folder and open it in the agent.
3. Paste the prompt below, press enter, and answer its questions as it builds.

The prompt

Build me a website chatbot to replace MagicChat. Requirements:

- Node.js 22 + TypeScript: a Fastify server. `npm start` serves http://localhost:4870: admin at /admin,
the chat API at POST /api/ask and the widget script at /widget.js.
- Sources: a sitemap URL or a list of URLs (fetched with `undici`, text pulled out with
`@mozilla/readability` and `jsdom`), PDFs (`pdf-parse`) and pasted text. Respect robots.txt and skip
pages whose content hash has not changed on a re-crawl.
- Storage: better-sqlite3 with the `sqlite-vec` extension in ./data/bot.db: sources, chunks (about 500
tokens with 50 overlap), embeddings, and every question and answer with its sources.
- Answering: embed the question, take the 6 closest chunks and call an OpenAI-compatible endpoint
(BASE_URL, API_KEY, CHAT_MODEL, EMBED_MODEL in .env). The system prompt allows answers only from the
chunks, cites source URLs, and says "I don't know, email <SUPPORT_EMAIL>" when nothing relevant comes back.
- Widget: a vanilla JS chat bubble (<script src=".../widget.js" data-bot="...">) with streamed answers
(server-sent events) and source links under each answer.
- Abuse limits, all in .env: per-IP rate limit, a daily message cap, a maximum question length, and CORS
restricted to ALLOWED_ORIGINS.
- Admin: one password in .env. Add sources, re-crawl now or every N days (`node-cron`), browse
conversations, export them as CSV and list the questions that got "I don't know".
- Single owner. No accounts, no telemetry. Outbound calls: your own sites and your LLM endpoint only.
- Out of scope: YouTube transcripts, several bots, Slack or WhatsApp channels, human handoff.
- README: serving it behind HTTPS, exactly what text is sent to the LLM, and a warning that anything
you index can be pulled out through the public endpoint.

Opens your agent with the prompt filled in, you press enter. Or copy it and paste it into an empty folder.

Share on X

Max MiskowSponsoredAI agents, automations and MVPs, built for your company.Visit site ›

Then keep going

Next prompts, same folder.

The first prompt builds the core. Run these one at a time in the same folder to grow it into your own MagicChat.

Extend the website chatbot in this folder. Requirements:

- After an "I don't know" answer, the widget offers a short form (name, email, message) that stores
  the lead in SQLite and emails it to SUPPORT_EMAIL through nodemailer.
- Add a /admin list of leads with the conversation attached, a status (new, answered) and CSV export.
- Validate the email, add a honeypot field and a per-IP limit of 5 submissions an hour.
- Never send the lead's data to the LLM endpoint.
- Update the README with the SMTP settings.

Add YouTube sources to the website chatbot in this folder. Requirements:

- Accept a video, playlist or channel URL in /admin; list videos with `yt-dlp` (JSON output) and fetch
  captions only, never the video files.
- Chunk each transcript by timestamp, store the video URL with a ?t= offset and show it as the source link.
- Skip videos without captions and list them in the admin page.
- Re-check each channel weekly for new videos.
- Update the README with installing yt-dlp.

Add a weekly quality report to the website chatbot in this folder. Requirements:

- A `node-cron` job (Mondays 08:00) groups last week's questions by embedding similarity and counts
  how many got "I don't know" or a thumbs-down from the widget.
- Add thumbs up and down buttons to the widget and store the votes.
- Email the owner the top 10 unanswered topics with example questions and the closest existing page.
- Keep the report as Markdown in ./reports/YYYY-MM-DD.md.
- Update the README.

What you lose

Scheduled re-indexing, so the bot follows your docs without you touching it
Training on YouTube videos and many file types out of the box
Ready-made install guides and integrations for Shopify, WordPress and WooCommerce
A live dashboard of activity and question counts, and webhooks
Someone else carrying uptime and abuse protection for a public endpoint

Why people still pay

Running a public endpoint that spends model tokens on strangers' questions, and keeping its answers current as the site changes, is a chore most site owners do not want. MagicChat turns it into a copy-and-paste embed for a few dozen dollars a month, which is also why non-developers pick it.

Moatexecution polish integrations

Do the math

MagicChat or your own copy?

Over

Your usage

Count my time3h ×$/h

MagicChat, 1 yr$348

Your own copy, 1 yr$80running + build session

You keep$269pays for itself in the first month

Estimate · A $5 VPS or a free tunnel, plus model tokens: a few cents per hundred answers on a small model. Indexing a small site costs well under a dollar once. The build is token usage on your agent plan.

Leaving MagicChat

Take your data, then cancel.

1 · Take your data out of MagicChat

1The CLI documents only login, create-bot, add-source, chat and ask: no command to list bots or export conversations or sources.
2Your knowledge sources are your own pages, PDFs and pasted text, so keep the originals; there is nothing to export there.
3Conversation logs: no export is described in what we read. Check the dashboard at app.magicchat.ai and ask support@magicchat.ai for a copy before you cancel.

2 · Cancel MagicChat

1Open the billing portal from your account and cancel there.
2Access continues until the end of the current billing period.
3Subscribed on iOS? Cancel in Apple ID settings at least 24 hours before renewal.

Paid fees are non-refundable except where the law requires it.
The Terms do not say what happens to your bot's data after you cancel: ask support to delete it if that matters.
Remove the widget snippet from your site when you leave, or visitors will see a dead chat bubble.
MagicChat's own guide notes the public API needs no key, only the bot's public ID, so don't index anything private.

Checked 2026-10-19 against MagicChat's docs: 1, 2, 3, 4

The escape hatch

Don't feel like building it? These are already free.

AnythingLLMOpen sourceMIT app for chatting with your documents on any model; the embeddable website chat widget is Docker only.★ 67k · active Oct 2026 · docker to self-host
WebVisit ↗OnyxOpen sourceAI chat and search over company knowledge with connectors; MIT outside its ee folders, which carry a separate licence.★ 33k · active Oct 2026 · docker to self-host
WebVisit ↗RAGFlowOpen sourceApache-2.0 retrieval engine with document parsing and an agent layer; heavier to host than a small crawler.★ 92k · active Oct 2026 · docker to self-host
WebVisit ↗

All 4 free alternatives to MagicChat ›

The numbers

MagicChat pricing

Plan	Monthly	Annual, per month	What you get
Free	$0	$0	1 chatbot, 100 messages a month, up to 50 pages, embed on your site, API access, community support.
Lite	$29	—	1 chatbot, up to 1,500 messages a month, up to 500 pages, 1 seat, manual refresh of sources.
Starter	$59	—	1 chatbot, up to 4,000 messages a month, up to 1,000 pages, 1 seat, manual refresh.
Growth	$129	—	Up to 2 chatbots, 12,000 messages, 10,000 pages, 4 seats, platform integrations, rate limiting, automatic monthly refresh.
Scale	$429	—	Up to 10 chatbots, 50,000 messages, 50,000 pages, 10 seats, webhooks, weekly refresh and a daily scan.
Enterprise	custom	—	Custom volume and limits, priority support, signed DPA or BAA, HIPAA eligible.

Free tier: 100 messages a month, 50 pages and one chatbot.
Billing: Monthly by default; paying yearly saves about 40% but the page shows no yearly prices. A 7-day free trial is listed on paid plans, though the page is inconsistent about which ones. Prices exclude tax. Cancel from the billing portal and keep access to the end of the period; paid fees are non-refundable except where the law requires.
Hidden costs: One visitor question plus the AI reply counts as one message, so a busy site hits the cap quickly. Extra 5,000 messages cost $59 a month and removing MagicChat branding costs $59 a month, each as an add-on on any plan.
Source: magicchat.ai

Also on the list

Chatbase$40/mo · Customer supportYES

Intercom$39/mo · Customer supportKINDA

Chatwoot Cloud$19/mo · Customer supportKINDA

Questions

MagicChat, briefly

Not answered here? Tell us and we'll add it.

Ask a question ›

Free covers one chatbot, 100 messages a month and up to 50 pages, with embed and API access.

Yes. An AI coding agent such as Claude Code, Codex or Cursor can build a usable personal version in one session with the prompt on this page, running on your own machine.

About $29 a month on the Lite plan, $348 a year (checked 2026-10-19).

Scheduled re-indexing, so the bot follows your docs without you touching it; Training on YouTube videos and many file types out of the box; Ready-made install guides and integrations for Shopify, WordPress and WooCommerce; A live dashboard of activity and question counts, and webhooks; Someone else carrying uptime and abuse protection for a public endpoint. If any of those matter to you, keep paying.