AI access · InterfacesREV. 2026-A

How AI assistants access this website

Besides the pages for people, this website has entry points for programs. Through them an AI assistant can have tools from /tools run calculations or look up our prices, instead of searching pages for figures. This page explains how to connect and where the limits are.

  • no sign-up
  • read-only
  • stores nothing permanently

As of

Three entry points, the same data as on the pages

MCP server

/mcp

For assistants that speak the Model Context Protocol. 54 tools, from company profile and prices to the calculators from /tools.

A2A endpoint

/a2a

For agents using the A2A protocol 1.0. Runs 50 tools from /tools, with the same calculation as /mcp.

Markdown

Accept: text/markdown

If a program asks for it, a page returns its text as Markdown. Browsers still get HTML.

CALL TOOLAI ASSISTANT/mcp OR /a2aCALCULATED ASON /toolsANSWER WITHSOURCEREAD PAGEAGENTPAGETEXT AS MARKDOWNCALL TOOLAI ASSISTANT/mcp OR /a2aCALCULATED AS ON /toolsANSWER WITH SOURCEREAD PAGEAGENTPAGETEXT AS MARKDOWN
FIG. 1 · TWO WAYS TO THE SAME DATAGOLD = ENTRY POINT FOR PROGRAMS

All three read from the same sources as the website. Company details, price list and the tools’ calculations are not maintained twice anywhere: if an hourly rate changes, it appears on /preise and in the MCP server’s answer at the same time after the next release.

Connecting the MCP server

Address
https://aureum-tech.com/mcp
Transport
Streamable HTTP, JSON-RPC 2.0 over POST
Sign-in
none
Access
read-only

An MCP client that connects servers by address over Streamable HTTP needs only this address. Your client’s documentation tells you exactly what the entry is called. Many expect a configuration like this:

{
  "mcpServers": {
    "aureum-website": {
      "url": "https://aureum-tech.com/mcp"
    }
  }
}

Some clients also want the transport stated, usually as “http” or “streamable-http”. No key or token is needed.

What happens when you connect

  1. initialize: The client registers without a session ID. The response carries Mcp-Session-Id in its header, and the client sends this ID along with every further request.
  2. notifications/initialized: The client confirms that the registration is complete.
  3. tools/list: The server lists all tools with their description and input schema.
  4. tools/call: The client calls a tool with its inputs. The result arrives as JSON in the text field of the response.

This is what registration looks like by hand, for instance to try it out in a terminal:

curl -i https://aureum-tech.com/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{
        "protocolVersion":"2025-06-18","capabilities":{},
        "clientInfo":{"name":"test","version":"1.0"}}}'

The response arrives as an event stream (text/event-stream), and the ID is in the header line mcp-session-id. Next, a tool call using this ID; the confirmation from step 2 (notifications/initialized) is left out here for brevity, but a real client sends it first:

curl https://aureum-tech.com/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -H "Mcp-Session-Id: <Kennung aus der ersten Antwort>" \
  -H "MCP-Protocol-Version: 2025-06-18" \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{
        "name":"preise_abfragen","arguments":{"zielgruppe":"firma"}}}'

If you open /mcp in a browser, it sends a GET and receives the message “Nur POST (JSON-RPC)”. That is the correct answer, not a fault. /mcp/health shows whether the server is running.

Available tools

Four tools answer questions about us:

  • firma_und_kontakt: company profile with address, phone, availability and register details
  • leistungen_auflisten: all service areas, each with its page
  • preise_abfragen: hourly rates, travel and surcharges from the same source as the price list
  • website_suchen: search across all pages of aureum-tech.com

In addition there are 50 of the 51 tools from /tools (all except the e-invoice check), each under a name like werkzeug_email_check. They calculate with the same logic as the page in the browser, and every answer names the tool’s page, the date of the data and the limits of the result. The inputs a tool expects are listed on its page in the section “For AI assistants”.

Limits that keep the service running for everyone

The server is public and has no user accounts. For that reason the cost per request is capped, and the throttles on tools that query other servers apply to all calls combined.

Limits of /mcp and /a2a
LimitValueWhat happens then
Size of a request900 KBHTTP 413
Open MCP sessions, all clients combined500HTTP 503 with Retry-After: 30
MCP session with no activity30 minutesThe session expires, the client receives 404 and registers again.
Response time per call of a tool from /tools without network lookup, including waiting time10 secondsAborted with an error message from the tool
E-mail check30 checks per minute, 300 new DNS checks per hourThe tool asks you to try again later.
Third-party, AI crawler and link preview checks20 checks per minute each, 150 new checks per hour eachas with the e-mail check
Gas calculator with live valuesFees fetched again at most once a minuteAnswer with the values of the last minute; without live values, with the fixed figures on the page and a note

A result that is still in the cache (five minutes per domain or address) does not count towards the hourly cap. If a tool itself reports an error (throttle, time limit, unusable input), it appears in German in the fehler field of the response. Inputs that do not match the schema are rejected by /a2a with the JSON-RPC error -32602 and by /mcp with an error response from the MCP SDK.

Calling tools through the A2A protocol

Endpoint
https://aureum-tech.com/a2a
Protocol
A2A 1.0 over JSON-RPC 2.0
Header
A2A-Version: 1.0
Method
SendMessage

Agents that speak the A2A protocol call the same 50 tools through /a2a. A message with a data part {"werkzeug": "…", "eingabe": {…}} runs a tool, and the response contains the same result as with /mcp:

curl https://aureum-tech.com/a2a \
  -H "Content-Type: application/json" \
  -H "A2A-Version: 1.0" \
  -d '{
    "jsonrpc": "2.0",
    "id": 1,
    "method": "SendMessage",
    "params": {
      "message": {
        "messageId": "9b2f6c1e-4d1a-4c3b-8e55-0a7d2f3b1c90",
        "role": "ROLE_USER",
        "parts": [{
          "data": {
            "werkzeug": "werkzeug_email_check",
            "eingabe": { "domain": "ihre-firma.de" }
          },
          "mediaType": "application/json"
        }]
      }
    }
  }'
  • A message with text only returns the list of tools with their input fields. If the text names a tool, you get its full details including example questions.
  • The agent works by fixed rules and does not understand free-form language.
  • It creates no tasks and stores nothing. Every answer is a direct message, so ListTasks returns an empty list.
  • It does not offer streaming or push notifications, and it does not accept files.
  • If the header A2A-Version: 1.0 is missing, it answers with the error -32009: the specification reads a missing value as version 0.3.

The four tools for company profile, services, prices and search exist only through /mcp. The same body size limit, time limit and throttles apply to /a2a. The description for programs is in the Agent Card.

The pages as Markdown

If you request a page with the header Accept: text/markdown, you get its text as Markdown instead of the HTML, without navigation and scripts. Browsers never send this header and see the page as usual. These versions are generated from the finished pages when the website is built, so they are as up to date as the pages themselves.

curl -H "Accept: text/markdown" https://aureum-tech.com/preise

The response carries Content-Type: text/markdown; charset=utf-8 and Vary: Accept. The exceptions are the guides overview and the articles from the content management system: both are filled only when requested, so they still return HTML. /llms-full.txt provides the text of these pages in a single file.

Where programs look up what is available

These files describe the entry points in formats that agents and checking programs look for on their own. The home page points to the API catalogue, the OpenAPI description and this page in the Link header of its response.

  • /.well-known/api-catalogAPI catalogue according to RFC 9727. For /mcp and /a2a it names the description, this page as documentation and, for /mcp, the status check.
  • /openapi.jsonOpenAPI 3.1: the HTTP interface of the MCP endpoint.
  • /.well-known/mcp/server-card.jsonMCP Server Card with address, transport and tool list of the MCP server.
  • /.well-known/agent-card.jsonA2A Agent Card: endpoint, protocol version and capabilities of the A2A agent.
  • /.well-known/agent-skills/index.jsonDirectory of Agent Skills. It contains two guides for agents: using the tools through MCP, and preparing a service enquiry for a person.
  • /llms.txtShort version of the website for language models, with the important pages.
  • /llms-full.txtThe text of the pages in a single file.
  • /auth.mdSign-in guide for agents. Mainly it says that there is no sign-in here.
  • /robots.txtRules for crawlers. All are allowed; the known AI crawlers are listed individually.

What happens to the inputs

A call through /mcp or /a2a sends the inputs to our server, which calculates with them and returns the result without storing inputs or result permanently. Only the results of network checks are kept in memory, together with the domain or address checked, for at most six minutes and with no link to the caller. As with every page request, only the request is logged, without its content, and what the assistant itself does with your details is governed by its provider (details in the privacy policy (Datenschutzerklärung)).

An access point like this for your own systems

What this website offers for its own data, we also build for businesses: an MCP server through which an AI assistant accesses your inventory management system or a database, with permissions you set beforehand. How we go about it is described on the page about MCP integrations.

View MCP integrations
+49 3744 4399760Mon–Fri 8:00–17:00
CallCallback

PARTS LIST · SHEET 1/1

What this page is made of

Values measured at build time. With JavaScript, your browser measures them again live.

ITEMPARTTYPESIZE
01index.htmlDocument≈ 29 kB
02styles.cssStylesheet≈ 20 kB
03scripts.jsScript≈ 13 kB
04archivo-variable.woff2Font≈ 58 kB
05ibm-plex-mono.woff2Font≈ 30 kB
06favicon.svgImage< 1 kB

Advertising or analytics cookies: none. Without your consent, the page loads nothing from third-party servers apart from Cloudflare, our delivery service. Fonts, scripts and images are hosted by us. That is how we build your IT, too.

OUTSIDE OFFICE HOURS

Feel free to call anyway

Our regular office hours are Monday to Friday, 8:00–17:00.

In the evening and at weekends the number is forwarded to us, and often someone will pick up. We cannot promise that, though.

If nobody answers, leave a short message on the voicemail: your name, your number and what it is about. We will call back as soon as we can, on the next working day at the latest.

The call costs nothing. If someone has to start work straight away outside office hours, the surcharge in the price list applies.

Call now+49 3744 4399760

Prefer to write? Use the contact form or e-mail [email protected]. We will read it on the next working day.

DIAGNOSIS · AT-K2

What do you need help with?

↵ SELECTESC CLOSECALL: +49 3744 4399760

Custom settings

Which services may load?

Strictly necessary

The choice you make here and your settings for colour scheme and motion. These are kept as small entries on your device and are never sent to a server. They are needed for the site to work and therefore cannot be switched off.