AI Suite MCP

Release v0.10.0

For details see https://github.com/autodudesde/ai-suite-mcp/releases/tag/v0.10.0

Important:

Please clear all caches after updating.

Requires AI Suite 14.6.1, 13.18.1 or 12.24.1.

Release Notes: v0.9.0

This release adds support for MCP revision 2026-07-28 and for TYPO3 installations without Composer, TCA-driven content search, EU AI Act provenance recording for every tool call, and backend links in tool results that open in the running backend session, plus many tool-readability and translation fixes.

For details see https://github.com/autodudesde/ai-suite-mcp/releases/tag/v0.9.0

Important: Please clear all caches after updating. Breaking: the four content-analysis audit tools are merged into auditContent, the folder batch tools into a folderIdentifiers property, and the search table settings are removed (searchable tables now resolve from the TCA).

Release Notes: v0.8.0

This release adds SEO, accessibility and content-analysis audits as MCP tools (question coverage, content gap, topic cluster and competitor comparison), free read access to stored audit results and credit accounting with a per-session spending cap, plus several token, transport and tool-schema fixes.

For details see https://github.com/autodudesde/ai-suite-mcp/releases/tag/v0.8.0

Important: Please clear all caches after updating.

Release v0.7.0

For details see https://github.com/autodudesde/ai-suite-mcp/releases/tag/v0.7.0

Important:

Please clear all caches after updating.

Breaking change: the write mode "auto" has been removed, as it was the only mode that could silently send a write to live. An installation still configured as "auto" now writes into a draft workspace; set "MCP Write Mode" to "Live" explicitly to keep live writes.

Release v0.6.0

For details see https://github.com/autodudesde/ai-suite-mcp/releases/tag/v0.6.0

Important:

Please clear all caches after updating.

`mcpWriteMode=workspace` no longer falls back to live writing. If an installation relies on that fallback, switch it to `mcpWriteMode=auto` or `mcpWriteMode=live`.

Release Notes: v0.5.0

Features:

- Workspace write mode is now the default: MCP writes land in a draft workspace, and a per-user workspace is created automatically if the editor has none. ext:workspaces is therefore a hard requirement instead of a suggestion, and the mode stays configurable (mcpWriteMode: workspace / auto / live).
- Reworked the tool surface into a consistent lexicon: read tools are named read*, listing tools list*. Renamed getPageContent, getPageTree, getFileInfo, getServerInfo, getRecordSchema, getFlexFormSchema, getTaskStatus, getTaskResults, findStaleContent, getTables, getPageTypes and getContentTypes accordingly.
- Added readContentTree to read every content element of a whole page subtree, with UIDs and page-wise pagination, in a single call instead of looping readPageContent per page.
- Added readEditorialGuidelines so tone, audience and style rules for a given page reach the model before it writes.
- Added readChildren for container and inline children, plus copyMediaReference and replaceMediaReference for image and media relations.
- Added the safe-edit tools replaceText, patchText and bulkReplaceText: small, targeted edits with old/new excerpts instead of rewriting a whole field.
- writeRecords now writes inline children in the same call, accepts common alias shapes (type for the real CType field), refuses a newly created container that no child points at, and offers an optional atomic mode that rolls the whole batch back on failure.
- previewRecords shows a per-field old/new diff on updates and accepts the same record shapes as writeRecords.
- batchGenerateMetadata can now target a whole page subtree via rootPageId + recursive instead of requiring a resolved UID list; the subtree expansion is capped at 50 pages and honours the caller's page permissions.
- Translation tools now offer the site's actual languages as an enum instead of a free-text ISO code, so an unavailable language is rejected by parameter validation. On single-language installations the translation tools are hidden entirely.
- tools/list now ships per-tool annotations, so MCP clients such as Claude Desktop can offer "Always allow" for read-only tools while write, AI and media tools keep their per-call confirmation.
- The initialize response carries a compact session orientation (sites, write mode, writable tables), so the model knows what it may write from the first turn.
- listContentTypes returns the page's valid colPos values, a FlexForm bucket with the matching readFlexFormSchema follow-up and a container's inline children; readRecordSchema reports readOnly, eval, relation kind and content kind.
- readRecords curates system and noise fields out of its output and resolves relation UIDs to readable titles, honouring the user's table access.

Bugfixes and improvements:

- Fixed an OAuth refresh-token rotation issue that logged connectors out mid-conversation: concurrent refreshes no longer race each other, a rotated token is separated from a revoked one, and a 30 second grace window (RFC 9700) serves a concurrent retry instead of raising a false token-theft alarm and revoking every session.
- Closed a critical authentication bypass in the OAuth consent step, where a client-supplied user ID could mint a full token without a password. The identity is now bound to an HMAC-signed consent ticket issued only after successful authentication. Added per-IP rate limiting on the token and registration endpoints and fixed same-origin validation, which previously rejected the consent POST on production setups without configured origins.
- FlexForm values are validated against the real data structure before they reach the DataHandler. A wrongly shaped value used to be stored as an unreadable FlexForm, which then broke every later read of the record and the frontend rendering, while the write itself still reported success.
- Records without a type value are now validated and described against their default sub-schema, instead of against every column of the table. Creating a plain EXT:news article no longer fails with required fields that belong to other record types.
- writeRecords no longer invites a JSON-string payload in its schema. Batches sent as an escaped string arrived truncated and were silently lost; a truncated payload is now recognised, reported with its length, and the model is told to split the batch.
- batchTranslatePage and batchTranslateFileMetadata now refuse an unresolvable target language. Previously it silently became the source language, and the batch paid per page to translate content into the language it was already in.
- Batch results report isError: true on partial and total failure, instead of always reporting success. Errors carry a machine-readable type, and DataHandler errors are translated into classified messages (e.g. missing permissions) rather than raw log output.
- The prompts capability is now announced over HTTP as well. A spec-compliant client never asked for a prompt before, so the whole prompt surface was unreachable outside stdio.
- Fixed a TypeError in resources/list with typed SDK request parameters.
- readPageContent now lists the other record tables stored on a page, so a record folder is no longer read as empty and the model stops falling back to a page or content element when a domain record was asked for.
- generateImage and uploadMedia are flagged destructive, since they write to live file storage that no workspace can undo.
- Trimmed tool descriptions and shared guideline snippets to cut the token cost of the tool catalogue on every turn.
- Fixed a uid_local resolution issue in the record access checks and several FlexForm record schema details.

Maintenance:

- Removed the tools generateContent, generatePageTree, generateLandingPage, generateMetadata, optimizeContent, auditContent, getColumnPositions and getOperatingGuidelines. Their results can be composed from the read and write tools, and each removed entry frees context on every turn. generateFileMetadata and batchGenerateMetadata stay, since they use vision models and server-side prompts a client cannot reconstruct.
- Dropped the OAuth scopes mcp:manage, mcp:glossary and mcp:easy-language. They were offered on the consent screen but gated nothing; the glossary is applied by the translate tools under mcp:translate.
- The operating guidelines no longer force the model into an approach decision before it can pick a tool and no longer ask it to wait for an approval the host is responsible for; they now steer toward sensible defaults, field discovery and correct container handling.

Important:

Please clear all caches after updating and compare database.

This release renames and removes MCP tools. Connected clients cache the tool list from their authorization, so reconnect or re-authorize the MCP connector after updating. ext:workspaces is now a hard requirement, and new MCP writes go into a draft workspace by default.

# Release Notes: v0.4.0

## Features:

* Added the `uploadMedia` tool to upload images and videos via URL, base64 or online media (new `mcp:media` scope)
* Added the `compareWithLive` tool to diff a workspace draft against the live version
* Added the `getFlexFormSchema` tool and a local stdio transport
* Added a `raw` flag to `readRecords` to return stored HTML/RTE markup
* Added batch support for the move and copy record tools
* Enforce TCA-required fields when creating records via MCP
* Surface the `ai_suite_mcp` version and warn clients on a stale token

## Bugfixes and improvements:

* Improved TYPO3 v12/v13 cross-version compatibility (edit-access checks, FlexForm schema, workspace comparison, v12 IRRE elements)
* Fixed type-aware RTE handling so bodytext is no longer flattened on write, with more accurate read hints
* Improved `readRecords` output (HTML-stripped hints judged on the raw value, each field rendered once, truncated results flagged)
* Return structured per-record batch errors with readable output and schema introspection across the record tools
* Fixed the MCP resource handler for TYPO3 v14 (replaced the removed `SiteLanguage::getTwoLetterIsoCode()`)
* Handle `DateTimeImmutable` values in `ReadRecordTool`

## Important:

Please clear all caches after updating and compare database.

Release Notes: v0.3.3

Bugfixes and improvements:

- Updated composer dependencies
- Added missing extension icon

First release of the AI Suite MCP