Skip to content

Coworker

Coworker, contentworker's assistant in the admin — opening it, how page edits go through it, which changes wait for your confirm, what it can read and change, and how long conversations are kept.

Coworker is contentworker’s assistant: a chat panel in the admin. Ask it about your content, how a feature works or what is waiting to be published, or have it create, tag, publish or schedule entries. It answers from live data through tools, and anything that changes something waits for your confirm.

It works as you. Every tool call runs with your live permissions, in the space and environment the header has chosen, so it can never do what you couldn’t do yourself or what an MCP client with your access couldn’t.

  • The Coworker button in the header, or ⌘J (Ctrl+J) anywhere, which also closes it.
  • Changing something from a page. See Edits from a page.

At lg width and up the panel is a column beside the page, and it reopens on your next visit if you left it open. Below that it is a sheet between the header and the tab bar; Esc closes the sheet.

Enter sends and Shift+Enter starts a new line. On a phone or tablet Return starts a new line and the send button sends. Stop ends a reply in progress. A message written while a reply is still streaming is held above the prompt, where you can remove it, and is sent when the reply ends.

Changes you make on the Content pages go through Coworker. The page’s button doesn’t change anything itself: Coworker runs the tool for exactly what you pressed, with no model involved. Each edit stays in the conversation as a line that says what it did and links to the entry, such as Published · Spring launch notes.

  • An action stops at a confirm card in the panel, which opens if it was closed, and the page waits for your answer. Confirm does it; Not now leaves everything as it was. Actions are publishing and unpublishing an entry, moving it to the trash, deleting it for good, restoring a version from the history panel, and changing several entries at once.
  • An edit runs at once, doesn’t open the panel, and leaves a line. Edits are saving a draft, creating an entry, duplicating one and restoring one from the trash.

When Coworker can’t be reached, the panel says so in one quiet line (Can’t reach Coworker right now. Edits you make on the page still go through.). The page then makes each edit itself, confirming in its own dialog where it confirmed before.

Ask in plain language: “What changed since it was last published?”, “Which of these are still drafts?”, “Suggest tags for this entry”. Reads run without asking and show as a line in the conversation.

Anything the model proposes that changes something stops at a confirm card, the same as a page action, and nothing is sent until you confirm. A declined card reads “left as it was”. A card that nobody answers is let go after a day.

A card that publishes, unpublishes or restores an entry is held to the entry as it was when the card was made: the version on your screen for a page’s card, the current version when the model proposed it. If someone saves a newer draft before you confirm, the card is refused and nothing changes, so a confirm never publishes or overwrites a draft you haven’t seen. Ask again for a card that shows the entry as it is now. Your role and access to the space are checked again for every message, so a role change or a revoked space takes effect in an open conversation.

The chip beside the send button says what “this” means: the entry open on the page, the content type whose list you are on, or Everything elsewhere. Open it to switch between the page and everything. The page’s context travels with each message.

Point at a message to show its actions:

  • Copy copies the text.
  • Edit on one of your messages opens it for changes. Send again replaces that message and every reply after it. Escape cancels.
  • Answer again on the latest reply asks for a new answer to the same question.

A message from before a confirm card offers only Copy: editing or asking again never reaches back past an action. If a reply fails, the panel says so with Try again.

New conversation, in the panel’s menu, asks first and then deletes the conversation. Changes you confirmed stay made and stay in the audit log.

It offers a fixed set of MCP tools, by their MCP names.

Kind Tools
Reads, no confirm docs_search, docs_read, model_list_content_types, model_get_content_type, entries_query, entries_get, entries_suggest, content_search, entries_list_versions, entries_diff_versions, entries_list_trashed, releases_list, releases_get, comments_list, tasks_list, taxonomy_list_tags, assets_list, workflow_task_list
Edits, run at once from a page entries_create, entries_update, entries_duplicate, entries_restore, entries_set_metadata, entry_apply_tags, comments_add, releases_add_entry
Actions, always ask first entries_publish, entries_unpublish, entries_trash, entries_purge, entries_restore_version, entries_bulk_action, schedule_action, releases_publish

A write the model proposes always waits for your confirm, whichever row it is in. Every other MCP tool stays MCP’s alone. Whatever arguments it is given, a call runs in the space and environment you are working in, with your role’s scopes: without permission to publish, a publish fails.

Replies come from Workers AI through AI Gateway. If the models error or don’t answer in time, a simple built-in fallback answers the turn with the same tools and the same confirms, and page edits never involve a model.

A conversation is yours alone: only you can open it. It is kept until it has been untouched for 30 days, then deleted; any new activity starts that period again. It is also deleted if your account is removed or disabled. Signing out ends your access to it until you sign back in.

Confirmed changes are ordinary changes with an ordinary audit trail, and outlive the conversation.