TYPO3 MCP: What Is the Model Context Protocol and How Do You Use It?

Ask a chat assistant for a meta description and the real work starts afterwards: copy, open the page in the TYPO3 backend, paste, save. Then the next page. The Model Context Protocol, MCP for short, takes that loop away. Your assistant reads the page tree, drafts pages and uploads files inside TYPO3 itself. It only gets the rights you choose to give it.

Anyone searching for TYPO3 MCP wants two answers. One is a definition. The other matters more: what a connected AI may touch. You will find both below, followed by a six-step setup and a security checklist.

Start with the server, which exposes data and actions. On a TYPO3 site it runs inside your installation. The host is the AI application you work in, such as Claude Desktop or Visual Studio Code. The host opens one client per server connection, and that client carries the conversation.

Servers offer three kinds of primitive, in the specification's terms:

For TYPO3, tools matter most. They are what changes your site.

Local development usually pipes messages through standard input and output (stdio) on one machine. Across the network, the assistant talks to TYPO3 over HTTP. For that route, the MCP architecture overview recommends OAuth, the common standard for delegated access. Both routes come ready for you in AI Foundation.

Behind the diagram are four hops:

1. An editor types a plain question.
2. The assistant fetches the server's tool list and calls the tool that fits.
3. TYPO3 runs the query as the connected backend user, within that user's rights.
4. The answer returns along the same path.

More than 100 MCP tools ship with the foundation and the AI extensions combined. Every one of them slots into the installation you already have.

For developers: your own tables, tools and skills

The same MCP server can open your own tables and endpoints to the assistant.

• Your own tables. Discover Extension Tables finds tables in your extensions. Each one you enable gets nine tools for creating, reading, updating and deleting records.
• Custom tools. Point a new tool at a PHP class, a REST endpoint or a webhook, all from one dialog.
• Prompt templates. Clients fetch them through prompts/list and prompts/get, including the built-in add_content_text_block for turning an editor's brief into a text element.
• Skills. Every AI extension ships a Markdown skill file for your assistant, and the Skill Hub imports more.

Six steps from now, an assistant you pick is working in your TYPO3 installation, but only in a draft workspace and under a user with limited rights. Bring TYPO3 v12, v13 or v14, PHP 8.2 or later, and Composer. Workspaces comes along as a dependency; the draft workspace itself you create once.

Step 1: Install AI Foundation

• Run composer require nitsan/ns-t3af in your project directory.
• Finish with vendor/bin/typo3 extension:setup and vendor/bin/typo3 cache:flush.
• You will not be asked for a licence key or a registration, on any domain.
• The code is GPL-2.0-or-later and public in the TYPO3 Extension Repository (TER).

Nothing else is needed.

Step 2: Switch on the MCP server

• Go to AI Foundation, then MCP Server, then the Advanced tab.
• Tick Enable MCP Server and clear the cache.
• Check that the module header now shows the status Online.
• Test from outside: request /mcp on your site without credentials and expect a 401.

A 401 is what you want here. No valid token or certificate, no entry.

Step 3: Give the assistant its own user

• Set up a backend user reserved for the assistant. Keep your admin account out of it.
• Configure the user's group under AI Permissions; the dialog takes you through modules, features, records, limits and a review.
• Limit its page tree and languages with TYPO3's usual mount point and language settings.
• In Limits, cap daily requests and turn on workspace enforcement.
• Keep publish rights and live-workspace access off this user.

The assistant now acts with this user's rights.

Step 4: Connect your assistant

• Open the Connect tab and pick Remote (OAuth), which the screen marks as recommended.
• Copy the server URL. It ends in /mcp.
• Open the setup guide for your client. The module carries guides for 15 clients, Claude Desktop, Cursor, VS Code, Windsurf, n8n and the MCP Inspector among them.
• Approve access in the browser, signed in as the new user. OAuth 2.1 with PKCE (Proof Key for Code Exchange) protects this step, so an intercepted sign-in code is worthless.

Once you have signed in, the client shows up under OAuth Tokens. Each entry carries its backend user, its expiry time and a revoke button of its own.

Working locally with the DDEV development environment? AI Foundation connects directly through the TYPO3 command line too. Building extensions with an assistant is covered in our guide to Claude Code and TYPO3.

Step 5: Get to know the tools and try them in the Playground

No surprises later: the MCP Tools module lists every tool the assistant can call.

• Tools & Resources groups them under the label "TYPO3 Core" (AI Foundation's own core set), AI extensions, your own tables and custom tools.
• The Playground runs any single tool against your installation. Pick a tool, fill in its parameters and run it. No client is needed.
• Next to it you see the tool's calls per week, its success rate and, where provided, example prompts.

The count you see depends on which AI extensions and tables you have enabled.

Step 6: Start in a draft with a first task

• Select a draft workspace in the MCP Server module header. Live is workspace 0; drafts start at 1.
• Begin with a read-only job, like finding pages without a meta description.
• Check the first write in the Workspaces module, then publish it yourself.

Letting the AI write straight to live is a team decision, not a technical one.

The screenshots show the AI Foundation interface, captured on t3planet.de.

The MCP server guide in the AI Foundation documentation covers every connection method in detail. For the first steps with someone at your side, book an MCP demo for TYPO3 with Jürgen.

This post is its own example. Claude, Anthropic's AI assistant, reaches our TYPO3 installation on t3planet.de through the AI Foundation MCP server. Here are two moments from the work, one reading and one writing.

Reading: is the topic already taken?

The first task came before any writing. Claude had to find out whether one of our pages already covered the topic. Two calls to the pages_search tool settled it:

1. Pages with "mcp" in the URL: 0 results.
2. Pages with "MCP" in the title: 2 results, our AI Foundation post in English and German.

Two read-only calls turned a hunch into a decision, and nothing on the site changed. This post explains the protocol; the existing one covers the product.

Writing: the page is built hidden

Writing uses the same connection. With the copy approved, Claude builds this page in both languages, title, description and content included, and leaves it hidden. A person on our team decides when it goes live.

The first post we built this way taught us one thing. The database reported a clean save, yet the rendered page had defects. A clean save proves nothing, so every page now gets a browser check in both languages, on desktop and phone. Open every AI-written page in a browser before it goes live. The check fits the review your team already runs.

Pick a mode, and you pick the author. AI Foundation's two MCP modes also decide where your content travels.

• Context mode (default). The assistant's own model writes, and MCP tools place the result in TYPO3. Connect Claude Desktop, for instance, and Anthropic's model is the author. Anything the assistant reads for the task is sent to Anthropic.
• Native mode. The text is generated inside TYPO3 by the providers configured in AI Foundation. Your assistant gives the instruction. Tool results it reads still reach its provider.

After a mode switch, reconnect the client so it gets the new tool descriptions.

Native mode is where multi-LLM in TYPO3 happens: several large language models side by side. Seven platforms are built in, each a dependency in the Packagist listing for version 1.2.1: OpenAI, Anthropic, Google Gemini, Mistral, Azure, OpenRouter and Ollama. Any OpenAI-compatible provider works too. Run Ollama and the model stays on your own infrastructure. On your own API keys, the default, content goes straight from your server to the provider you picked, with T3Planet outside that data path.

Which provider sees which content is among the first questions in any data protection review. With AI Foundation that choice stays yours, and the software fits around it. Our post on the MCP-First AI Foundation for TYPO3 goes deeper into managing providers, prompts and budgets.