Setup Guide

Connect to the MCP Server

Your knowledge base, callable as tools by anything that speaks the Model Context Protocol. One endpoint, one header, nothing to install.

Before You Start

Two Things

πŸ”‘

An API key

Created in the admin under API Keys. The key decides which site the client can see, so there is no site identifier to pass and no way to reach another tenant's content.

🧩

A client that speaks MCP

Anything implementing the Model Context Protocol. If you would rather call it directly, it is JSON-RPC over HTTP and you need no library at all.

Setup

Four Steps

1

Create a key for the site

In the admin, open API Keys, choose the client and the site, and generate the key. Copy it now: it is shown once.

2

Decide what the key may do

Leave the scope as query for a client that only reads. Add ingest only if the client should also write to the knowledge base, because add_knowledge, ingest_url and ingest_sitemap refuse a key without it.

Do not reuse the widget key. That one sits in the public markup of your site where anyone can read it, so it carries query only, by design.

3

Point the client at the endpoint

The address is https://api.slatech.ai/mcp, and the key goes in the X-Api-Key header. A missing or invalid key is refused with 401 before any tool runs.

4

Ask what it can do

Call tools/list. The server answers with every tool, its description and a JSON-Schema for its arguments, so a client configures itself. Discovery is part of the protocol, so there is no separate document that can drift out of date.

Reference

The Tools

πŸ”

query_knowledge_base

Ask a question and get an answer grounded in the site's indexed content. Counts against query quota, exactly like a question asked on your website.

πŸ“„

list_documents

List what is indexed for the site, with the status and chunk count of each document. Read-only and not metered.

✍️

add_knowledge

Add a question and answer pair directly. Useful when the answer lives in someone's head rather than on a page. Needs ingest.

🌐

ingest_url

Fetch and index one page. Private and loopback addresses are refused, so the tool cannot be aimed at something internal. Needs ingest.

πŸ—ΊοΈ

ingest_sitemap

Index every page an XML sitemap lists, sitemap indexes included. This is how you bring a whole site in with one call. Needs ingest.

Good to Know

Limits and Metering

πŸ“Š

Counted on their own meters

MCP calls are recorded separately from chat traffic, so you can see what the integration costs without it hiding inside your website numbers.

⏱️

120 calls a minute

Per key. An agent works in bursts, so the ceiling is set well above a polling dashboard. Exceeding it returns 429 with Retry-After.

🏒

One key, one site

No tool takes a site identifier. The client and site come from the key alone, so a caller cannot reach another tenant by changing an argument: there is no argument to change.

i

Using the older manifest API?

GET /v1/mcp/manifest and its execute endpoint still work, and integrations built on them keep running. That one is a plain HTTP shape rather than the protocol, and its operations take a site identifier as an argument. New work is better pointed at /mcp.