AI Agent 

Purpose 

The AI Agent is a docked right-edge assistant in the TYPO3 backend. Editors ask in plain language, review proposed changes as cards, and apply them only when they confirm. Reads run automatically; writes never go live without approval.

Path (settings): AI Foundation > AI Agent

Open the panel: toolbar Ask AI Agent, or Ctrl/Cmd+Shift+K (optional Ctrl/Cmd+K in the agent preferences stored in the browser).

Prerequisites 

How to use it 

  1. Open a backend module (Page, List, Filelist, …). The agent uses that module and page (or folder) as context.
  2. Open the panel with Ask AI Agent or the keyboard shortcut.

  3. Ask in the editor's backend language, or use:

    • /tool_name — run a known tool by name
    • @table:uid — attach a record to the request
    • starter chips — common tasks for the current module
  4. Read tools answer in plain words. Write tools show a preview card first.
  5. Review the card (fields, target workspace), then Execute or Decline. Applied changes can be undone from the card when the agent offers Undo. A stale card is not written over newer edits.

Conversations 

Conversations are stored on the server for each backend user. Opening the agent resumes the last-viewed conversation for the configured scope (page, module, or user). Use the conversations list in the panel header to switch, rename, or remove conversations. The AI provider is fixed for a conversation after the first answer.

Workspace 

Confirmed changes follow the workspace selected under AI Foundation > MCP Server (same preference as the MCP server). Live is allowed unless the editor's group has workspace enforcement that blocks it. The panel shows which workspace a change will go to before you apply.

Permissions and logs 

The agent only offers tools and fields the editor may use: page tree, tables, fields, files, and workspaces follow the editor's own backend rights and AI Permissions. A refusal explains why in plain words. Every refused, hidden, or failed request is written to AI Usage & Logs.

Credits 

In T3Planet Credits mode, the agent calls the chat completions endpoint and shows remaining credits in the panel. Exhausted credits lock the composer until you top up. See T3Planet Credits.

Settings administrators usually change 

Configure these under AI Foundation > AI Agent:

  • agentConversationScope — which conversation opens automatically (page / module / user)
  • agentSessionListEnabled — show the conversations list
  • agentSessionListDefaultFilter — initial list filter
  • agentMaxSessionsPerScope / agentMaxSessionsPerUser — retention caps (0 = unlimited)
  • agentConversationRetentionDays — inactive conversations → trash
  • agentHistoryTokenBudget — how much earlier chat is replayed to the model
  • agentContinueAfterConfirm — continue a multi-step plan after Execute / Decline
  • agentMaxReadToolsPerTurn / agentMaxWriteDraftsPerTurn — per-turn budgets

Cleanup of trashed conversations and related turn rows: t3af:agent:conversations:cleanup.

Limits 

  • Batch MCP tools records_apply and records_undo are not offered in the AI Agent (one-record approval cards). External MCP clients still use them; see Batch writes.
  • New content elements get a default column when the plan omits one, so layout constraints (for example content_defender) can resolve the column on save.