Your AI Assistant Can Run Your Website Chatbot Now
Your AI Assistant Can Run Your Website Chatbot Now
I shipped a new product this week: chatsnip — a chatbot for your landing page that you install by pasting one script tag, and that your AI assistant operates for you. No dashboard. No settings maze. You never log into anything.
This post is about why it exists and the one design decision that makes it different from the fifty other website chatbots you've seen.
The problem with every chatbot SaaS
If you run a micro-SaaS, you already know the pitch: visitors land on your page with questions, most leave without asking them, and a chatbot catches some of those conversations. The pitch is fine. The products are not, for two reasons:
1. The dashboard tax. Every chatbot product comes with an admin panel you have to learn, configure, and revisit. You wanted answers on your landing page; you got another SaaS to babysit.
2. The token markup. Most of them resell LLM completions at 5–20x cost, wrapped in "message credits". Your bill scales with your success, and the vendor's margin lives inside your conversation volume.
The bet: agents are the new admin panel
Here's the observation chatsnip is built on: in 2026, the people buying tools like this already have an AI assistant with a terminal — Claude Code, Codex, whatever. That assistant is better at operating software than any dashboard you could build.
So chatsnip has no dashboard at all. After checkout you get a prompt — literally a block of text you paste to your assistant. The assistant installs a small CLI, interviews you about your product ("what does it do, what are the plans, what tone"), configures the bot, sets your spend limits, and hands you the script tag. Three minutes, and the human never touches a config screen.
Ongoing operations work the same way. Once a week you ask your assistant:
./chatsnip hosted digest -since 7d
and get back something like: 37 conversations this week — 5 technical support, 1 complaint (handled, contact captured), and a lead named Sam who left sam@dev.io asking for a call about the Pro plan. The bot is trained to do the most valuable thing a landing-page chatbot can do: when someone shows buying intent or frustration, it offers a personal follow-up from the founder and asks for a name and contact. Those get extracted automatically.
Your keys, your costs
chatsnip doesn't resell tokens. On both plans, completions run on your own OpenRouter key — encrypted at rest on the hosted plan, plain env var if you self-host. You pick any model on the market, you see your real LLM spend on your own bill, and you set the budgets: tokens per day, tokens per visitor, messages per minute. If your key runs dry, visitors get a polite "taking a short break" instead of an error, and your assistant sees the problem on its next status check.
This also means I have no incentive to inflate your conversation volume — my price is flat.
Small enough to not matter
The widget is ~6 KB of dependency-free JavaScript — no iframe, no framework, no cookie banner. It remembers visitors in their own browser (localStorage), repaints transcripts instantly, and long conversations get compacted server-side so context never breaks mid-chat.
The entire backend — HTTP server, SQLite persistence, LLM client, billing, CLI — is one native binary of about 240 KB, written in machin, the language I've been building for AI agents. That's not a flex (okay, it's a small flex); it matters because the self-host story becomes trivial: one file, one command, runs on a €3 VPS.
Pricing
Self-host: $79 one-time (launch price). You get the binary, a personalized install command, and every update for a year through the same URL. Run unlimited chatbots on your own server forever.
Hosted: €19/month, with a 7-day free trial. I run the daemon, storage, and digests; you bring your OpenRouter key. Five distinct chatbots/sites are included — your whole indie portfolio — and each extra site is €2/month. The same bot on more of your own pages (your blog, your docs) is free: origins aren't billed.
Try it
The fastest way to understand chatsnip is to watch your own assistant set it up. Start the trial at chatsnip.intrane.fr, paste the prompt it gives you into Claude (or any assistant with a terminal), and have a chatbot answering on your landing page before your coffee cools.
And if you're reading this on blog.intrane.fr — the chat bubble in the corner is chatsnip, eating its own dog food.