TYPO3 MCP prompts: ten starters and the Playground

Good first TYPO3 MCP prompts only read: they ask about your pages, languages, redirects and files, and change nothing. Below are ten starters in plain words, each tested on our demo install, with the tools the assistant used and what it found. MCP, the Model Context Protocol, lets an assistant such as Claude work inside TYPO3 through tools that AI Foundation provides for free. You will also meet the Playground, where you run one of those tools yourself, without any assistant. Last comes the one habit to keep before your first change.

A good first prompt reads, names the place and asks for evidence. Five habits make that easy:

• Read first. Start with questions. Your assistant acts as the TYPO3 backend user who signed in, so a request to change something is carried out at once.
• Name the place. Say which site, page or folder you mean. Our demo runs two sites, and a German page has a different language ID on each.
• Use your own words. Write as you would to a colleague: the old TX-6 model, not a table or field name. The assistant finds the records.
• Ask for the evidence. Ask which pages, elements or records the answer comes from, so you can open them in the backend.
• One job per prompt. One question, then the next. Short prompts give answers you can check.

We ran these ten prompts on our demo install on 5 October 2026, as an administrator working on Live. Each one only read. The demo runs two fictional brands: Mustermann Outdoor, which sells e-bikes, tents and jackets, and Stadtwerke Musterstadt, a utility.

All ten use the core tools of AI Foundation, which are free. The answers are summed up here; your assistant words them in its own way.

Three answers show what a starter gives you.

The old product name. Asked where the TX-6 still appears, the assistant searched the text of all content elements and found three in each language. Two are fine: the TX-7 page says the new model replaces the TX-6, and the manuals page keeps the old manual for owners. One is outdated: the Spring Sale deals still sell the TX-6. News records sit in their own table, so ask for them separately. Changing a name everywhere is the second post of episode 3.

The redirects. The assistant listed all nine and named four to check. One is a chain of two hops, one uses 302 for a permanent move, one is disabled, and one points to a PDF. It did not fix anything. That is the right order: you decide which ones change.

Your own rights. The assistant reported that it works as an administrator with every right on every table, language and file. That tells you how careful to be: with these rights, a change is live as soon as it is saved. Who can edit what, and what your assistant can do, is episode 12 on 8 January 2027.

The Playground runs one MCP tool at a time, straight from the backend, with no assistant and no chat. Open AI Foundation, then MCP Tools, then the Playground tab:

1. Pick a tool in the list, such as pages_get.
2. Fill in its parameters. The form comes from the tool's own description, so each field shows its type, such as uid (int).
3. Select Execute and read the answer: the raw data your assistant would receive.

The panel next to the form lists Example Prompts for the selected tool, such as Show me the page with UID 1 including its translations. They show the kind of sentence that makes an assistant choose this tool, so they are a good source of starters of your own.

Use the Playground to see what a tool returns before you ask an assistant, or to check an answer you did not expect. One caution: it runs tools for real, as your backend user. A tool that writes, such as content_delete or write_table, changes your TYPO3 here too. Start with the tools whose names end in _get, _list or _search.

Starter 10 is the first half of a habit. A single field change is written at once (a dry run and records_undo are for records_apply batches, shown in episode 3), so every change runs in four steps:

1. Read. Ask for the current value, as starter 10 does, and keep it.
2. Change. One field, one language at a time.
3. Check. Open the page, or ask the assistant to read the value again.
4. Put back. If you do not want the change, ask the assistant to write the old value back.

Where you want a review step first, AI Foundation can hold changes to pages and content in a TYPO3 draft workspace. The second post of this episode shows that route from chat: draft, review, publish and discard. Files, file metadata, redirects, scheduler tasks and cache clears act live even in a workspace.

The video for this episode runs one starter in Claude on our demo install, then opens the Playground and runs the same tool by hand. Shown in Claude; other MCP assistants work too. Nothing is changed.

Every starter here uses AI Foundation's core tools, which are free and open source, no licence key, no registration. AI Assistant, AI Accessibility, AI Chatbot and AI Search add their own MCP tools for writing, translation, alt text, chatbots and search. Each needs its own licence. Together with AI Foundation they bring 100+ MCP tools.

What your assistant reads, it sends to its vendor, as with any chat. Keep that in mind before you ask about pages that are not public yet.

Once these starters feel familiar, the series goes further. Episode 3 makes everyday edits from chat and changes a product name everywhere. Episode 4 asks your TYPO3 fifteen questions and checks your SEO by conversation. Next week, episode 6 clones a campaign page with all its subpages. Not set up yet? TYPO3 MCP setup takes four steps, and TYPO3 MCP clients compared helps you pick an assistant.

The series comes from T3Planet by NITSAN Technologies, a TYPO3 Association Gold Member with more than 15 years of TYPO3 work. Its episodes are written by Jürgen, our AI Consultant for TYPO3. Every episode runs its job on our demo install before it is published.

All episodes of MCP in Practice for TYPO3