Frequently asked questions about AI Suite MCP

AI Suite MCP is a TYPO3 extension that provides a server according to the Model Context Protocol. This allows an AI client such as Claude Desktop, Claude.ai, ChatGPT, Claude Code or Open WebUI to connect directly to your backend and work there with the same permissions as the linked backend user. More on this in the Introduction.

The place where work is done. You operate the AI Suite backend module in the TYPO3 backend, while AI Suite MCP brings the same functions into the chat of your AI client. Additionally, there is the structural part that does not exist in the backend module, such as reading page trees, creating, moving, or deleting records, and searching for content.

Yes. AI Suite MCP is a standalone extension but relies on the AI Suite and is installed together with it. The feature and model permissions, as well as the connection to external AI providers, come from the AI Suite. Without the AI Suite, AI Suite MCP will not function.

Especially when work is distributed across many pages or records. Typical cases include translating entire page trees, adding missing metadata, recurring text corrections across many elements, and creating page structures. For a single text correction, the backend is usually faster.

Not by default. The preset write mode workspace saves every change in a draft workspace that must be published before it appears on the website. If you want it differently, you can switch to live in the Configuration. There are two exceptions: uploadMedia and generateImage create physical files and write live in every mode.

TYPO3 14.3.x/13.4.x/12.4.x (at least 13.4.1 or 12.4.11) and PHP 8.1 or newer (PHP 8.2 from TYPO3 13). In addition, the AI Suite in the version matching the TYPO3 major version, since AI Suite MCP 0.10.0 that is 14.6.1/13.18.1/12.24.1 and newer. The TYPO3 system extensions Workspaces and Reports are installed as mandatory dependencies, Workspaces because the default write mode is based on it, Reports for the security check.

Yes, since version 0.9.0. The extension ships the MCP SDK itself and therefore also runs in TYPO3 installations without Composer (classic mode) under TYPO3 14.x/13.x/12.x. See Installation.

For the AI-supported tools, yes; for the rest, no. Generation, server-side translation, image creation, batch processes and audits run through the AI Suite and require a valid key with the appropriate license package. Reading, searching, creating and editing records, text replacements, creating translation records, translations by your client's model, and embedding media run solely through TYPO3 and work without a key.

Credits are consumed by the AI functions, the batch processes and the three checking audit tools (auditSeo, auditAccessibility, auditContent). The remaining reading tools including readAuditResults, all record editing, localizeRecord, translateRecord and translatePage without model as well as uploadMedia cost nothing. The complete breakdown is listed under Available Tools.

With batchGenerateMetadata and batchTranslatePage across a page tree, every page costs credits, and a subtree is quickly larger than expected. That is why a rootPageId is limited to 50 pages. If there are more, the call is rejected before anything is billed, and it states the number of pages found. An explicit list of page UIDs remains unlimited, because it was compiled deliberately.

The tool remains visible in the tool list, the model calls it, and the call ends with a license error. This is the difference from a missing feature permission, as a missing permission removes the tool from the list beforehand. If a tool remains visible and fails only upon calling, check the license and not the backend group.

Billing runs through the AI Suite and is independent of whether a request comes from the backend module or from an MCP client. So there is no separate quota for AI Suite MCP. For packages and prices, see Packages/Licenses.

The endpoint requires HTTPS in production, authenticates via OAuth 2.1 with PKCE or via a static token, limits every request to 1 MB by default (mcpMaxRequestBodyMb) and allows 100 requests per minute per token. With every request, it is checked whether the backend user is still active, and a password change revokes all of their tokens. Details under Security & Operation.

No. A client only sees tools for which the linked backend user has permissions, and normal TYPO3 checks on pages, tables, and fields apply during the call. An editor with access to two page trees can edit exactly those two page trees via AI Suite MCP. If you want to exclude certain tables fundamentally, even for administrators, use mcpExcludedTables.

With the tools that call an external AI model, the content to be processed goes to the provider configured in the AI Suite for the respective function. The same providers, models and data paths apply as in the backend module, see Models & APIs. In addition, there is the client itself. A chat service running on third-party infrastructure sees everything the model reads through the tools.

AI Suite MCP writes two files of its own. var/log/aisuite_mcp.log receives the complete trace with OAuth events, client IP, user UID and excerpts of the tool arguments and responses (up to 500 characters each), which regularly contain prompts and content excerpts. var/log/aisuite_mcp_warnings.log contains only warnings and errors. Tokens, hashes and email addresses are masked before writing. The detailed log can be switched off via mcpLogVerbose. You define the retention period and deletion yourself.

AI Suite MCP processes the same data as the AI Suite and does not add any new providers, so the data protection assessment of the AI Suite remains unchanged. Two points are added. The AI client used is a separate recipient and must be considered, and the detailed MCP log contains personal data and requires a defined retention period.

In the backend module “AI Suite” in the MCP tab via “Revoke”. This takes effect immediately and regardless of whether the client is still connected. Alternatively, you remove the connector on the client side. A password change of the backend user also revokes all of their tokens.

Yes. The status report contains the entry AI Suite MCP Security under System → Reports (TYPO3 14: Administration → Reports) and reports typical misconfigurations there, for example allowed HTTP or empty allowlists in the production context.

Yes, since version 0.9.0. Every write by an MCP tool is recorded as AI-generated in the transparency register of the AI Suite, with the feature “MCP server” and the client used. Read-only calls leave no entry. The records appear in the “AI content” area of the AI Suite and carry the usual marking in the backend, see AI Content (EU AI Act).

Guides are available for Claude Desktop, Claude.ai, ChatGPT, Claude Code, the MCP Inspector and Open WebUI. In principle, every client that speaks Streamable HTTP according to the MCP standard works. An overview with login methods and reachability is available under Connecting Clients.

That depends on the client. Claude.ai, ChatGPT and Open WebUI run on third-party infrastructure and call your server from outside, so they need public HTTPS with a certificate from a recognized CA. Claude Code, the MCP Inspector and Claude Desktop via mcp-remote or the local stdio transport run locally and also reach internal hosts and *.ddev.site, self-signed certificates only if the operating system trusts the CA. A custom connector in Claude Desktop, on the other hand, needs a server publicly reachable via HTTPS, just like Claude.ai.

OAuth is the standard way and the only one for externally hosted clients. A static token is set up faster and is suitable for local clients, CI, and scripts, but it expires after mcpTokenLifetimeDays and must then be renewed. OAuth tokens renew themselves.

Yes. Every token is tied to exactly one backend user, and several users can connect their own clients at the same time. Each of them sees only their own permissions and page trees. A user can also have several clients connected in parallel, with at most 20 active tokens. If another one is issued, the server revokes the oldest one, and the client using it loses access.

Yes, with a local client. For DDEV, the CA must be in the operating system's certificate store, which mkcert -install takes care of. If TYPO3 runs on the same machine as the client, it can also work completely without HTTP, tokens, and OAuth via the Local stdio transport.

In most cases, yes. Custom connectors are not automatically active in every conversation. With Claude.ai and Open WebUI, you need to enable the connector under the input field per chat, in ChatGPT you call the plugin in the chat via @ (see ChatGPT). Otherwise, the model will not see the tools and will claim it has no access.

Up to version 0.5.0, the initialize handshake failed for clients that send params._meta. As a result, the Claude.ai connector was never established in the first place. Version 0.6.0 fixes this, an update to 0.6.0 is enough. If the client still does not connect afterwards, the troubleshooting under Connecting Clients helps further.

The tool list depends on the AI Suite feature permissions of the backend group. If a permission is missing, the corresponding tool disappears from the list, without an error message. If enable_mcp_access is missing, the server rejects every request with a permission error. On single-language installations, the translation tools are also missing. Which permission unlocks which tool is listed under Permissions & Scopes. After adding permissions, the client must be authenticated again.

This is the normal case in the default write mode workspace. Changes are stored as drafts in a draft workspace and only appear after publishing. In the backend, you can see them as soon as you switch to the same workspace. With compareWithLive, a draft can be checked field by field against the live state.

Yes. The command ai-suite-mcp:cleanup should run at least hourly in production, via the scheduler or cron. Without it, the table of authorization codes grows without limit, session and task directories fill up, and expired tokens remain. See CLI Commands.

First in var/log/aisuite_mcp.log (provided mcpLogVerbose is on). If there is nothing there, the request did not reach AI Suite MCP or was rejected because of missing HTTPS (400) or a body that is too large (413), and the access log of the web server is the right source. A 404 points to a site prefix in the URL, repeated 401 to a discarded Authorization header. The most common cases are collected under Connecting Clients, client-specific ones on the respective subpage.

If your installation is behind HTTP Basic Auth, this collides with AI Suite MCP, because both use the same Authorization header. The MCP paths and the OAuth discovery paths therefore have to be excluded from Basic Auth. How this looks under Apache and Nginx is described under Production Operation.

Not on every one. The extension is designed for PHP-FPM and mod_php, that is one process per request. It must not run on runtimes with persistent workers such as FrankenPHP or RoadRunner, because otherwise read permissions can leak between backend users. With several application nodes, you need sticky sessions or a shared file system for the session and task directories.

Up to version 0.5.0, the write mode workspace fell back to a live write if no draft workspace could be provided. Since version 0.6.0, the call aborts with an error message instead, so that nothing ends up live unnoticed. Make sure that the affected backend user can get a workspace, or switch the write mode (mcpWriteMode) to “Live”. Also clear all caches after the update.

The mode auto was removed in version 0.7.0. It was the only mode in which a write could silently go to live, because the target depended on the user's workspace assignment. If auto is still in your configuration, AI Suite MCP now writes into a draft workspace. If you want to keep writing directly to live, explicitly set the write mode to “Live”. Only “Draft workspace” and “Live” are still available.

Version 0.9.0 merged and removed tools. auditQuestions, auditContentGap, auditTopicCluster and auditCompetitors are merged into auditContent with the type parameter, batchGenerateFolderMetadata and batchTranslateFolderMetadata into the folderIdentifiers parameter of the file batch tools, and replaceText into patchText. Update the tool names in your own prompts, skills and integrations. The mapping is listed in the “Breaking changes in 0.9.0” box under Available Tools.

Up to version 0.8.0, a link clicked in an external chat reached the backend without the editor's session cookie and therefore ended on the login screen. Since version 0.9.0, the links go through an intermediate page under /aisuite-mcp/open and open in the running backend session. Updating to 0.9.0 is enough, it is controlled by mcpBackendLinkBounce (default: on), see Configuration.

Possibly. Since version 0.9.0, searchContent searches every table with text fields that can be reached through the page tree. The former settings mcpSearchAdditionalTables and mcpExcludeAdditionalTablesFromSearch are ignored. If a table should stay out, add it to mcpExcludedTables. Personal data, the ChEddi chat history and configuration and system tables are already excluded by a shipped default list, see Security & Operation.