Agent readiness

TYPO3 agent readiness

What the TYPO3 stack offers AI agents today, and who provides each part.

An agent that works with a website needs content it can read and an interface it may call. It also needs rights that limit what it does, and a review step before anything goes live. This page lists what TYPO3 Core, official add-ons, Netresearch extensions and other community extensions provide for each of these needs.

This page leaves out capabilities that are documented but not released yet, or marks them as work in progress.

Status as of

Who provides what

Each table row names one or more of five origins. Netresearch's open-source extensions are also part of the TYPO3 ecosystem. The origins show who delivers which part of a solution.

TYPO3 Core
Core and the system extensions shipped with it. Some must be activated and configured.
Official add-on
Offerings of the TYPO3 project or the TYPO3 Company that are separate from Core, such as the TYPO3 Suite or the Dev Companion.
Netresearch extension
Open-source extensions developed by Netresearch, such as nr-llm, Cowriter or nr-browser-ai.
Community extension
Extensions by other agencies and developers, credited by name.
Netresearch integration
Selecting, connecting and configuring the parts for a concrete installation: permissions, models, data sources, approvals, quality checks and operation.

What TYPO3 Core brings as a foundation

Core provides the structures that a controlled agent needs: typed records, permissions, drafts and integration points. The LLM, MCP, WebMCP and A2A functions come from the extensions in the sections below.

Capability Origin and component What it provides
Structured records, fields and relations TYPO3 Core TCA and DataHandler

TCA describes tables, fields and relations. Every write to a TCA table goes through DataHandler, which evaluates the acting user's write permissions.

Sources: TYPO3 Explained: DataHandler

Users, groups and fine-grained rights TYPO3 Core Backend access control

Backend groups limit which tables a user can read or modify and which page-tree branches and file mounts the user can reach. They also limit which languages the user can edit. An agent that acts as a backend user has the same limits.

Sources: TYPO3 Explained: access control

Drafts, versioning and preview TYPO3 Core System extension workspaces

Changes stay as versioned drafts that editors can preview. Drafts pass review stages before publication. An agent can prepare a draft, and review and publication stay in the editorial process.

Sources: Workspaces manual

Background processing and queues TYPO3 Core Symfony Messenger integration

TYPO3 integrates the Symfony Messenger message bus. Messages are handled synchronously by default.

Limits: Asynchronous processing needs routing to a transport (Core ships a Doctrine transport) and a running messenger:consume worker.

Sources: TYPO3 Explained: message bus

Incoming reactions and outgoing webhooks TYPO3 Core System extensions reactions and webhooks

Reactions receive webhooks from external systems. Webhooks notify external systems when TYPO3 events occur.

Sources: TYPO3 Explained: webhooks and reactions

Structured form definitions TYPO3 Core System extension form

Each form is stored as a structured definition of elements, options, validators and finishers. nr-browser-ai derives browser tools from these definitions (see below).

Sources: Form manual

Pluggable translation workflow TYPO3 Core Localization handlers, TYPO3 14

TYPO3 14 introduced a modernized translation workflow in which translation services can be plugged in as localization handlers.

Limits: The handler API is still marked internal in TYPO3 14.3 and may change.

Sources: Changelog: Feature #108049

Making content readable for LLMs and external agents

llms.txt, Markdown versions of pages, structured data and JSON output give agents a representation of the site they can use directly.

Capability Origin and component What it provides
Provide llms.txt and agent skills from configuration Netresearch extension nr-wellknown (beta)

At deploy time, writes /llms.txt as a static file from a file or inline text in the site configuration. When skills are configured, it also writes /.well-known/agent-skills.json.

Limits: There is one file per docroot: in a multi-site instance, the configuration of the first site applies. The agent-skills.json format is not a widely supported standard.

Sources: nr-wellknown on GitHub

Generate llms.txt from the page tree Community extension web-vision/ai-llms-txt by web-vision

Builds llms.txt from the page tree, with configurable depth and page-type exclusions. It leaves out hidden, access-restricted and no-index pages. It also serves pages as Markdown and exports them through a CLI command.

Sources: ai-llms-txt manual

llms.txt with Markdown endpoints Community extension rtfirst/llms-txt by Roland Tfirst

Generates llms.txt with Markdown access to page content and can protect these endpoints with an API key.

Sources: llms-txt manual

Pages as Markdown Community extensionNetresearch integration b13/ai-bots-love-markdown by b13, used in the Netresearch demo

Serves a page as Markdown when the request sends Accept: text/markdown or uses a URL suffix. The Markdown carries the page metadata as front matter, and the HTML page links to it with rel="alternate".

Sources: ai-bots-love-markdown on GitHub

Describe content semantically Community extension brotkrueml/schema by Chris Müller

Adds schema.org structured data to pages as JSON-LD, so content types and relations are machine-readable.

Sources: schema manual

Pages and content as JSON Community extension friendsoftypo3/headless by the TYPO3 Headless initiative

Renders pages, content elements, menus and metadata as JSON, with multilingual handling and workspace preview.

Sources: headless manual

Deliver content by context Netresearch extension contexts, combinable with Cowriter

Shows or hides pages and content by configurable contexts. One context matches requests from known AI crawlers. Cowriter can create audience variants that an editor can bind to such a context.

Limits: The client sets the User-Agent header. The AI-crawler context therefore classifies requests and does not authenticate them.

Sources: contexts on GitHub · Cowriter on GitHub

  • The Netresearch TYPO3 demo serves a hand-curated llms.txt directly from nginx and offers every page as Markdown via b13/ai-bots-love-markdown. It does not use nr-wellknown.
  • llms.txt is a proposal for giving agents a curated overview of a site. It does not guarantee that AI systems use or rank the content. Markdown and structured data are representations that agents can use directly.

External agents can work with TYPO3

Tools in the browser, tools in the CMS and delegated tasks are separate capabilities, each with its own interface and limits.

Capability Origin and component What it provides
Local assistant in the browser Netresearch extension nr-browser-ai (beta, proof of concept)

Answers questions about the open page with Gemini Nano through Chrome's built-in Prompt API, on the visitor's device.

Limits: Runs only where the browser offers the Prompt API. The project evaluated WebLLM and did not adopt it.

Sources: nr-browser-ai on GitHub

Form-based tools for browser agents Netresearch extensionTYPO3 Core nr-browser-ai on top of the form system extension

Generates a JSON Schema from a form definition. Where the browser supports WebMCP, it registers the form as a tool with the browser's model context. The tool fills the visible form and queries the data source configured for the form.

Limits: It does not submit the form, run finishers or execute anything on the server. For a multi-step form, nr-browser-ai describes only the first page.

Sources: Decision record: forms as tools · WebMCP binding source

TYPO3 as an MCP server Community extension hn/typo3-mcp-server by hauptsache.net

Gives MCP clients access to the page tree and to workspace-capable tables, including their schema. Clients connect to an HTTP endpoint protected by OAuth. The server queues record changes in a workspace.

Limits: It uses the user's first writable workspace. Uploaded files land in storage immediately, because files are not workspace-versioned.

Sources: typo3-mcp-server on GitHub

Use external MCP tools inside TYPO3 Netresearch extension MCP client in nr-llm

Administrators connect external MCP servers. Imported tools start disabled. An administrator enables them one by one, and by default each call needs a human approval.

Limits: The client speaks HTTP (no stdio) and passes on text results only.

Sources: nr-llm manual: MCP servers

Address TYPO3 as an A2A agent Netresearch extension A2A server in nr-llm

TYPO3 answers A2A 1.0 requests over JSON-RPC and publishes an Agent Card with the skills of administrator-defined profiles. Registered clients authenticate with a token. They can follow task status, answer follow-up questions, cancel tasks and fetch results. Each client's tasks run as a dedicated non-admin backend user in a configured workspace.

Limits: Files are off by default. When files are on, a task accepts files from the client. To return a FAL file, the profile must also approve its clients as recipients, and an administrator must enable a tool that ships disabled. The task sends only files that the client's backend user may read.

Sources: nr-llm manual: A2A server

Delegate work to other A2A agents Netresearch extension A2A client in nr-llm

Administrators register remote A2A agents. The skills on their Agent Card become tools. These tools are off by default, and by default a run needs an approval before it delegates.

Limits: The client exchanges text only. While the backend chat nr-mcp-agent is installed, nr-llm does not offer delegation tools to agent runs yet (see the limits section).

Sources: nr-llm manual: remote agents

Personal agents with the editor's own rights Netresearch extension A2A tokens in nr-llm

Where an administrator allows it, editors issue their own revocable A2A tokens. A token is limited to permitted profiles and, by default, to a draft workspace and an expiry date. nr-llm checks every request against the editor's current groups.

Sources: nr-llm manual: A2A tokens

Official API for records and files Official add-on TYPO3 Suite API Hub

A REST API for records and files with OAuth 2.0 or static tokens. Each app is limited to its backend user's rights and to the actions it is allowed.

Limits: Part of the TYPO3 Suite, a commercial offering of the TYPO3 Company. The TYPO3 Company has announced an MCP server for the Suite, without a launch date.

Sources: TYPO3 Suite on typo3.com

  • The Netresearch TYPO3 demo has answered A2A requests since 17 November 2026. Its agent card is public. Tasks need a token that an administrator issues, and writes go to a draft workspace after approval.

Agents and workflows inside TYPO3

nr-llm runs agents within TYPO3 itself: with CMS context, tools, follow-up questions and approvals.

Capability Origin and component What it provides
Persistent agent runs Netresearch extension Agent runtime in nr-llm

Runs are stored, can run in a background worker and keep a timeline of their tool calls.

Sources: nr-llm manual: agent runs

Pause, ask and resume Netresearch extension Agent runtime in nr-llm

A run can pause for an approval or a typed answer and then continue, be rejected or be cancelled.

Sources: nr-llm manual: approvals and input

Read TYPO3 context Netresearch extension Read tools in nr-llm

Built-in read tools cover content, page tree, TCA and FlexForm schemas, TypoScript and TSconfig, files and diagnostics. Each tool and each group can be switched off.

Sources: nr-llm manual: read tools

Change content under control Netresearch extensionTYPO3 Core Write tools in nr-llm, DataHandler and workspaces

Writing tools ship disabled, and every call pauses for a human approval. Tools that write records go through DataHandler into the run's workspace.

Limits: Not every writing tool uses a workspace: cache clearing and the redirect tools are admin-only, and the redirect tools write live.

Sources: nr-llm manual: tools and workspaces

Reusable skills Netresearch extension Skill management in nr-llm

SKILL.md instructions are versioned and approved. A configuration loads them always or on demand. nr-llm never executes scripts that a skill references.

Sources: nr-llm manual: skills

Central rules for language and style Netresearch extension Prompt snippets in nr-llm

Personas, tone, audience and style are kept as reusable building blocks that extensions combine by tag.

Sources: nr-llm manual: prompt snippets

Knowledge from existing search systems Netresearch extension Retrieval tool in nr-llm

Returns cited site content from the installed search index (EXT:solr, ke_search or indexed_search) or from a database fallback.

Sources: nr-llm README

Work through many records Netresearch extension Run batches in nr-llm

An action can run over a folder or a page tree. Each proposed change waits on its own approval card.

Sources: nr-llm manual: run batches

Structured AI decisions Netresearch extension DecisionService in nr-llm

Returns typed yes/no, choice or score decisions. Thresholds and consequences stay in the calling code.

Sources: nr-llm manual: DecisionService

Structured and multimodal output Netresearch extension Shared services in nr-llm

Schema-validated JSON output on every supported provider, plus shared services for translation, vision, image generation, embeddings and speech.

Sources: nr-llm README

Applications built on this foundation

Netresearch extensions that use nr-llm for concrete editorial and site tasks, plus the integration of community AI extensions.

Capability Origin and component What it provides
Writing assistance in the editor Netresearch extension Cowriter

A CKEditor assistant that improves, summarizes, extends, corrects and translates text. It suggests page fields, including SEO and social texts, and generates alt texts. Optional features generate images and audience variants.

Sources: Cowriter on GitHub

AI chat in the backend Netresearch extension nr-mcp-agent (proof of concept)

A backend chat on nr-llm's agent runtime. It keeps conversations, uses nr-llm tools, runs work in the background and offers guided flows, for example for missing alt texts.

Sources: nr-mcp-agent on GitHub

AI search and site chat Netresearch extension nr-ai-search (beta, not publicly available, on request)

Answers visitor questions only from the site's own indexed pages. Each search or chat element can be limited to one page subtree.

Limits: The extension is not publicly available. Netresearch provides it on request. The Netresearch TYPO3 demo documents it as an extension advisor for the Extensions section. That chat is not publicly accessible, so anonymous visitors cannot start paid model calls.

Sources: Demo README: extension advisor

Content repurposing Netresearch extension nr-repurpose (alpha)

Turns a web page or PDF into one content brief. From the brief, it creates formats such as a podcast, a diagram, a story carousel, a summary, an FAQ, social posts, a newsletter, a slide deck and a handout.

Limits: nr-repurpose can hand approved social posts to an automation tool through a webhook, optionally signed. It sends no other format by webhook.

Sources: nr-repurpose on GitHub

Landing pages from a briefing Netresearch extension nr-landingpage (beta)

A backend wizard. The editor picks a template, answers briefing questions and reviews the generated page fields and content sections. Then the editor chooses where in the page tree the wizard creates the page. The template limits which content elements the wizard may generate.

Sources: nr-landingpage on GitHub

Bring community AI extensions under shared control Netresearch extensionCommunity extension nr-llm-compat (beta)

Routes the LLM calls of supported extensions through nr-llm and its policies. Supported extensions include AI SEO Helper, T3AI, AI File Metadata, Texter, Exception Solver and Sitescore.

Limits: Each integration is off by default and activates only for a supported version whose code contract still matches.

Sources: nr-llm-compat on GitHub

Automated translation Community extensionNetresearch extension Autotranslate by Thiele & Klose with nr-llm

Autotranslate translates pages and content, including scheduled batch jobs. With nr-llm as its translation backend it can use DeepL or a language model, with usage tracking and budgets.

Limits: Autotranslate 3.6 and earlier need an upstream change applied as a patch.

Sources: nr-llm manual: Autotranslate · Autotranslate on GitHub

A curated combination of the stack Netresearch extension TYPO3 AI Edition

A Composer metapackage that installs Netresearch AI extensions in versions tested together. CI installs it and opens each extension's backend modules.

Limits: The edition pins the versions that passed its compatibility run. Newer single releases, including nr-llm's A2A features, are not part of it yet. The run does not test model answers.

Sources: TYPO3 AI Edition on GitHub

  • Other agencies offer their own AI foundations. b13 AiM is a central AI layer that selects providers and models, enforces rules and tracks costs. NITSAN's T3AF is an open-source foundation with a built-in MCP server, provider management, budgets and permissions.

What keeps the stack controllable

Which tasks a site can hand to an agent depends on how the task is limited, how its result is checked and who decides on publication.

Capability Origin and component What it provides
Central model and provider management Netresearch extension nr-llm

Shared configuration for providers and models with criteria-based routing, fallback chains and local models via Ollama.

Sources: nr-llm manual: routing · nr-llm manual: fallback chain

Budgets and usage limits Netresearch extension nr-llm

Per-user budgets cap requests, tokens and cost per day and per month.

Sources: nr-llm manual: user budgets

Limit the scope of a task Netresearch extensionTYPO3 Core Assignment scope in nr-llm

A run can be bound to pages, records, tables and languages. DataHandler writes outside that scope are refused, for administrators too.

Sources: nr-llm manual: assignment scope

Drafts and human approval Netresearch extensionTYPO3 Core Editorial profile in nr-llm with Core workspaces

A configuration can require that every change is written as a draft in a workspace, administrators included, and that a second person approves.

Sources: nr-llm manual: editorial profile

Rules and a second assessment Netresearch extension Rule set judge in nr-llm

An optional second model checks finished texts against written rules and either reports or blocks violations.

Limits: It is off by default. What it catches depends on the configured rules and the model.

Sources: nr-llm manual: rule set judge

Repeatable quality checks Netresearch extension Quality evaluation in nr-llm

Golden prompt sets measure answer quality and flag regressions after changes to prompts, models or configuration.

Sources: nr-llm manual: quality evaluation

Mark AI-edited content Netresearch extensionCommunity extension AI-written marker in nr-llm, plus optional b13/ai-label by b13

Pages and content elements written by a run are marked as AI-written. b13's ai-label adds AI-origin labels, a review record and a visible marker for visitors.

Limits: A technical approval is not an editorial review of the content.

Sources: nr-llm manual: AI-written marker · ai-label on GitHub

Retention of run data Netresearch extension Governance policy in nr-llm

How long conversations, agent runs, approvals and other data are kept is configurable per data category.

Sources: nr-llm manual: governance policy

Manage credentials Netresearch extension nr-vault

Stores secrets with envelope encryption, per-secret access by backend group, a tamper-evident audit log and rotation of secrets and the master key.

Sources: nr-vault on GitHub

Protect outgoing HTTP calls Netresearch extension nr-http-guard (alpha)

Adds SSRF protection to TYPO3's RequestFactory. It blocks private, loopback and metadata destinations, checks DNS results and redirects, and pins connections to verified addresses.

Limits: Direct cURL calls and other HTTP clients are not covered.

Sources: nr-http-guard on GitHub

TYPO3 for coding agents

Agent readiness also covers development: coding agents that build and maintain TYPO3 projects need version-specific knowledge and runtime context.

Capability Origin and component What it provides
Version-specific TYPO3 knowledge Official add-on TYPO3 Dev Companion

A local MCP server for coding agents. It provides version-specific TYPO3 knowledge and read-only context from the local installation, and it ships task skills for common development work.

Limits: Experimental 0.x release. It writes nothing into the installation.

Sources: TYPO3 news: Dev Companion · Dev Companion on GitHub

Machine-readable documentation Official add-on docs.typo3.org rendering

Every manual on docs.typo3.org gets an llms.txt. Depending on the render, a manual also gets Markdown pages and JSON indexes.

Limits: This applies to the documentation platform, not to TYPO3 websites.

Sources: render-guides: llms.txt

Runtime context for development assistants Community extension typo3-ai-mate by Konrad Michalik

A development-only bridge to Symfony AI Mate. Its diagnostic tools report TYPO3's resolved runtime state, such as TCA, TypoScript, Fluid, dependency injection and events. Nearly all of these tools only read.

Limits: The tools run on demand through a CLI command. There is no long-running server process.

Sources: typo3-ai-mate on GitHub

Where the edges are today

This page does not count merged but unreleased work as available.

  • The chain from the backend chat to a remote agent and back is not available yet. While nr-mcp-agent is installed, nr-llm offers agent runs no delegation tools. nr-mcp-agent already contains a display for a step that waits for a remote agent. That display stays unreachable until nr-llm lifts this block. Keeping the final answer of such a step is in progress in nr-llm.
  • The A2A client exchanges text only. The A2A server accepts and returns files only when files are enabled.
  • nr-browser-ai is a proof of concept. Its form tools fill a form and query its data source. They never submit the form.
  • nr-mcp-agent is published as a proof of concept. nr-repurpose and nr-http-guard are alpha releases.
  • llms.txt, Markdown and JSON output make content usable for agents. They do not guarantee that an AI system uses or ranks it.
  • Automated checks of the AI Edition install the extensions and open their modules. They do not judge model answers. An approval in a workspace is a technical step, not an editorial review.