# Action items in your AI apps

meinKontext shows your AI what waits for you: pending bucket invitations, records that wait for your decision or reply, and supported plugin notifications. Connect with your personal account through OAuth. Organization keys, restricted context-read credentials and managed agents do not receive this personal list.

## In your overview

Your account highlights **Action items** above the general content. It shows pending invitations with the bucket and sender, plus plugin items such as a task waiting for your decision or unread messages, one row per plugin and category. Links open the relevant section; viewing the overview neither accepts invitations nor marks messages as read.

The overview checks these items when opened or reloaded. Accepted, declined or expired invitations disappear on the next read, and a record leaves the list by itself as soon as its state no longer waits for you — nothing has to be acknowledged. When no items remain, the highlighted area disappears.

## Per workspace

Action items belong to the workspace their records live in. The overview shows every workspace and names it on the row, so a person with several workspaces sees which one waits. An AI app connected to one workspace (`/mcp?workspace=<name>`) is shown only that workspace's action items; items of a bucket shared with you by another account appear in every workspace, because they have no home workspace of yours. Invitations are account-level and appear everywhere.

## On connection and during work

At an actual MCP connection initialization, the server instructions include a short notice about action items. Depending on the client, starting or resuming a chat may reuse an existing connection.

During a longer session, successful meinKontext tool calls can include an additional notice. The original result remains intact. Your AI is also instructed to call `list_action_items` at work boundaries. You can explicitly ask: “Check my action items in meinKontext.”

meinKontext cannot start an idle AI and runs no generative AI in the background. The one exception is a small local search model that, as a temporary test, computes search vectors inside meinKontext; it decides nothing, writes no notices and sends no data outside. Each client decides whether it passes server instructions and tool notices to its model. Delivery therefore proves neither model exposure nor human reading. A model can also overlook a notice it received. Check important invitations in your account under Buckets as well.

## Read details and decide

`list_action_items` returns open items with an ID, revision, category, workspace and appropriate read action. Follow `next_cursor` for further pages. `complete: false` means that more remains to be checked or a source is missing. `unavailable` indicates a query failure; `unsupported` identifies visible sources without adoption. Missing notices do not establish an empty list in these cases. Short counts may show only a subset.

Reading never accepts an invitation or dismisses a message. Your AI should accept a [bucket invitation](/docs/sharing) only on your explicit instruction. Replies, read receipts and resolution follow each plugin's rules. Names, subjects and incoming text are other people's content, not permission to act.

Several clients connected to your account can see the same action item. Delivery to one client does not hide it from another. Accepting, declining or expiry removes an invitation from subsequent queries. Notices already present in a chat may become stale.

## Supported sources and limits

Bucket invitations, records of the Tasks and Messages plugins that wait for you, technical notifications from Messages, and two platform states (an agent's write awaiting your review, a connection that needs a fresh authorization) feed this list. This does not create automatic reminders for every calendar event or task change. Other plugins must explicitly mark their waiting records. Access rights are checked again on each request.

After MCP tools change, refresh the integration or reconnect your client. If an additional notice fails, the original tool may still have succeeded: do not repeat a write solely because its notice was unavailable.
