Skip to content

An MCP server so AI agents can search your website

Appik Studio

AI agents do more than search the web. Claude, ChatGPT, Cursor and in-house agents call external tools through the Model Context Protocol (MCP). Every Askolia project can now publish an MCP server. A connected agent searches your published pages, documents and FAQ answers, reads the full source and cites its URL.

Product details in this article are as of November 1, 2026.

Why an agent uses your MCP server

An agent that browses a website loads whole pages, guesses their structure and sometimes stops at the first page it finds. With the MCP server, it calls a search that already knows your content. It receives short results with a title, an excerpt and the page URL, then opens the source it needs. It uses less context and cites your pages instead of a summary published elsewhere.

The server uses the index of your Askolia search, which covers pages crawled on your domain, additional websites, published documents and FAQ answers in their languages. There is no extra content to prepare.

Two read-only tools

  • search looks through the published sources. It accepts a query of 2 to 200 characters, an optional FAQ language (en, fr, de, es or it) and returns up to 10 results. Each result has an id, a title, an excerpt and, for a web page, its URL.
  • fetch returns the full text of one result from its id. Text longer than 20,000 characters is truncated.

These names match the tools ChatGPT expects from a search connector, described in OpenAI's MCP documentation. Both tools are declared read-only. The server writes nothing, sends no message and opens no conversation.

What the agent can read

The server exposes only what search already shows your visitors, which is indexed public pages, published documents and FAQ answers. Internal assistant instructions, conversations, reply drafts and visitor data stay out of reach. Pages from a removed additional website also leave the results.

The server instructions describe returned content as untrusted data and tell the agent never to follow instructions found in a source.

Turn the server on

  1. In the dashboard, open the project settings, then the "Search your site with AI" section.
  2. Under "AI agents (MCP)", click "Allow AI agents to search". Only the workspace owner can change this setting.
  3. Copy the address shown under "MCP server URL".

The address looks like https://tacit-goose-489.eu-west-1.convex.site/mcp/<public key>. The public key is your widget key, already visible in your pages' code. It gives access to public content only. The server is off by default. When you turn it off, the address returns 404.

Connect a client

Claude Code adds the server with one command:

claude mcp add --transport http my-shop https://tacit-goose-489.eu-west-1.convex.site/mcp/YOUR_KEY

In Cursor, add it to the project's .cursor/mcp.json file:

{
  "mcpServers": {
    "my-shop": { "url": "https://tacit-goose-489.eu-west-1.convex.site/mcp/YOUR_KEY" }
  }
}

In VS Code, add it to .vscode/mcp.json:

{
  "servers": {
    "my-shop": { "type": "http", "url": "https://tacit-goose-489.eu-west-1.convex.site/mcp/YOUR_KEY" }
  }
}

In ChatGPT and Claude, add a custom connector with the same address and no authentication, if your plan allows it.

The server uses the Streamable HTTP transport without sessions. Each request stands alone. CORS headers also let a web client call it.

Limits

  • The server runs keyword search and generates no AI answer. The agent writes its answer with its own model, so these calls use no AI search generations.
  • A project accepts 120 requests per minute across all clients. Beyond that, the server returns 429 with a 60-second retry delay.
  • Search indexing limits apply: 500 pages and documents per project, and only readable public HTML. PDFs are not imported.
  • The user must add the server to their client. No widely adopted mechanism yet lets an assistant discover a website's MCP server on its own.
  • The dashboard does not show agent requests.

Who it is for

A software company makes its documentation and FAQ available in Cursor or Claude Code, where its developer customers write code. A shop or service business gives its sales and support teams an agent that drafts answers from the official pages. An agency publishes a separate server for each client project.

The MCP server is available on every plan because it uses no AI generation. Human visitors keep the AI search built into your site, and the AI search overview explains how it works.