How to add an AI chatbot to a website with an AI agent (MCP)

The short version

Clara is an AI conversion agent that sits on a website, answers visitors in text (and out loud on the Pro plan), and captures leads or books appointments. An AI agent can set her up for a site owner in five calls. The sequence, in the order the server expects it:

create_trial_tenant (free 14-day trial, no card, every Pro feature; needs the site owner's real email) -> refresh_site_knowledge (or set_site_knowledge if the site isn't live yet) -> update_tenant_settings / configure_integration -> install the embed snippet -> get_upgrade_link to keep Clara after the trial.

Everything below is what each call needs and returns. Everything is free for 14 days with every Pro feature, voice included, and no card is taken.

Connect the MCP server

The server is a public Streamable HTTP endpoint at https://claraconverts.com/mcp. It needs no API key to connect, and it is stateless: each call is one request and one JSON response.

• Claude Code: claude mcp add --transport http clara https://claraconverts.com/mcp
• Claude.ai: add it as a custom connector using the same URL.
• Other clients and builders: any client that supports remote Streamable HTTP MCP servers, including custom connectors in app builders such as Lovable.

It is listed in the official MCP Registry as com.claraconverts/clara. Machine-readable entry points are at /agents.txt and /agents.json.

Step 1: create the trial

Before calling anything, ask the site owner for two things: the site URL and their real, monitored email address. The email is the only place their dashboard login link and trial reminders go. Never invent one and never reuse an address found elsewhere. Only one live trial is allowed per email.

Then call create_trial_tenant with owner_email and site_url. You can also pass display_name and conversion_goal (for example, what Clara should drive: a booking, a lead, a demo). It returns:

• a working embed snippet;
• a secret_key for every later step (see authentication);
• the trial's expiry date.

The owner is emailed a welcome message with a dashboard login link at the same moment, so they learn the trial exists without having to know an agent set it up. If that email cannot be delivered, the trial is cancelled and the next call tells you why. Ask the owner for a working address and try again.

If the owner is comparing options first, get_pricing and list_integrations are open to anyone and need no key.

Step 2: teach Clara the business

A new trial skips the site crawl so the first call returns quickly. Until you load knowledge, Clara works but knows nothing about the business, which makes a poor first impression. Do one of these:

• refresh_site_knowledge: crawls the live site and replaces what Clara knows. It can take up to a minute on a large site.
• set_site_knowledge: for a site that is not published yet (localhost, a preview URL, still being built). Pass each page's text, markdown or HTML, up to 300 pages. Once the site is live, run refresh_site_knowledge to replace it with a real crawl.

If the crawl cannot reach the site, the error says so and points you at set_site_knowledge.

Step 3: set her up

update_tenant_settings changes only the fields you pass: business name, conversion goal, brand colour, widget position, language, voice, the email captured leads go to, an outbound webhook, and the assistant's name and avatar. get_tenant_settings reads the current state back, which is useful to confirm a change took or to resume later.

To connect a booking, CRM, commerce or other system, call list_integrations to see what exists and what each one needs, then configure_integration. Some connect with an API key you pass in. Others return a link a human must approve. For Cal.com, call list_calcom_event_types first, because the event type has to be picked from the account's real list.

Step 4: install the snippet

get_embed_snippet returns the script tag, and the same snippet came back from create_trial_tenant. Place it once, site-wide. If you are working in the site's own code, put it in the root layout. The install guide has the placement for Next.js, Astro, Nuxt, SvelteKit, WordPress, Webflow, Squarespace and plain HTML.

The snippet's key is unique to this trial. There is no generic snippet, and a placeholder will not work, so use the one the server returned. The call also returns an optional agents.txt snippet. Publish it at the root of the owner's own site so other agents visiting that site can reach Clara's tools too.

Step 5: after the trial

The trial runs for 14 days. No card is on file, so nothing is charged automatically. When it ends, the widget stops responding on the site. The owner gets a reminder email three days before and another on the last day, each with a checkout link.

To keep Clara, the owner can pay from the Billing page of their dashboard, or you can call get_upgrade_link and give them the checkout link it returns. Either way, the same trial account is upgraded in place, so the embed snippet on the site keeps working with no change. Pro is the default plan, and Standard (text only) is available.

Authentication

Connecting and the three open tools (get_pricing, list_integrations, create_trial_tenant) need no credentials. create_trial_tenant returns a secret_key, and every later call uses it, either way:

• as an Authorization: Bearer <secret_key> header, or
• as a secret_key argument on the tool call, for clients that cannot attach a header partway through a session. Claude.ai's connector is one of them.

The header wins when both are present. The dashboard (leads, conversations, analytics) is intentionally for the human owner only. No tool here reads it.

Questions fréquentes

How can an AI agent add a chatbot to a website?

By calling ClaraConverts' MCP server at https://claraconverts.com/mcp. The agent calls create_trial_tenant with the site owner's real email and the site URL, refresh_site_knowledge so Clara learns the business, update_tenant_settings for name, goal and branding, then get_embed_snippet for the one-line script tag to paste on the site. The trial is free for 14 days with no card.

Do I need an API key to connect the MCP server?

No. Connecting, and calling get_pricing, list_integrations and create_trial_tenant, needs nothing. create_trial_tenant returns a secret_key that the later steps use, sent as an Authorization: Bearer header, or passed as a secret_key argument for clients that cannot attach a header mid-session.

What does the agent need to ask the site owner first?

Two things: the site's URL, and the owner's real, monitored email address. The email is the only place their dashboard login link and trial reminders go, so the agent must ask for it and never invent one. One live trial is allowed per email.

Why does Clara know nothing about the business right after the trial is created?

The trial skips the site crawl at creation so the call returns quickly. Run refresh_site_knowledge once the site is live, or set_site_knowledge with the pages' content if the site is still being built. Until one of them runs, Clara works but cannot answer questions about the business.

What happens when the trial ends?

After 14 days the widget stops responding on the site. No card is on file, so nothing is charged automatically. The owner gets reminder emails before it ends, and can keep Clara from the dashboard or with the checkout link get_upgrade_link returns. The embed snippet does not change when they upgrade.

Which AI agents and MCP clients can use it?

Any MCP client that supports remote Streamable HTTP servers. In Claude Code, run claude mcp add --transport http clara https://claraconverts.com/mcp. Claude.ai accepts it as a custom connector, and builders such as Lovable can add it as a custom MCP connector. The server is also listed in the official MCP Registry as com.claraconverts/clara.