Add an AI Agent to Your Website
The end-to-end journey — agentify your site with Edge MCP, then put the Omniagent on top with the Web SDK
You have a website. This guide puts an AI agent on it — one that doesn't just answer questions, but operates the site on the visitor's behalf: searches your catalog, fills the cart, prepares the booking, while the page updates in front of them. It takes two stages — first make your site's own operations available to an agent, then put the agent on top:
Stage 1 — Agentify the website
Expose the operations an agent may run and the state it may watch, using Edge MCP — an open integration on the WebMCP standard. Your AI coding tool does the work; you decide what gets exposed.
This stage needs access to your site's source code, because agentifying a site means adding code to it — an app you build and own, in whatever stack you use.
If your site is built on a hosted site builder or CMS and you only have the visual editor, you can't agentify it today — we don't ship integrations for those platforms yet. Stage 2 works either way.
Stage 2 — Add the AI agent
Put the Omniagent on the site: face, voice, and a live session. On an agentified page it discovers your tools automatically — no glue code.
This stage does not need source access. If your platform lets you paste a JavaScript snippet into the page — as most site builders and CMSes do — that's enough to run the agent. It just won't be able to operate the site until Stage 1 is done.
Multi-page site? Turn on persistence
On a traditional multi-page site every navigation reloads the page — and kills the session. Persistence keeps one conversation alive as the visitor moves through the site. Single-page apps don't need it — client-side navigation never reloads the page, so the session survives on its own.
NapsterCompanionApiSdk.init(token, {
persistence: { enabled: true },
});The result
A visitor talks to your website. The agent answers with your product knowledge, and acts: it searches, navigates, fills, and books — through the same functions your UI uses, as the signed-in user, within exactly the limits you approved.
This site is a working example. These docs are agentified with Edge MCP: the "Ask JC" agent in the corner calls this site's own tools to search the documentation and open pages for you, and it watches a live resource for the page you're on — so it always answers about what's in front of you. Ask it to find something and watch the page move.