Ask "what are my tasks today" while looking at a board, and chat used to answer from somewhere else entirely. It could not see your screen, had no idea what a task was, and owned no tool that could read a board. All three are fixed.
Chat knows what is on your screen
The open screen travels with your question. Which view you are in, which board, whether you are in the Bin, the search term and tag filters you have applied, the note you have open, and the cards you have selected.
Chips under the conversation show exactly what chat can see — Board view, This note, 3 selected — so context is never a guess.
Suggested questions follow the screen. On a board they become What are my tasks today?, What is in progress?, and What did I finish this week?
"This note", "these", "here" all resolve without naming anything.
Reading a board
get_boardreads a whole board with its cards grouped by column, and reports a total per column. It works on the workspace board and on any note's own board.Tasks are defined. There is no separate to-do entity in Notez, so chat treats the cards on the board you are looking at as your tasks — and a column whose name means finished (Done, Completed, Archived) as not pending.
Ask AI is where the notes are. A button on the board selection bar asks about the cards you have selected, and one in the note header asks about the open note.
One listing tool, with real filters
filter_notesis gone, folded intolist_notes. Every filter ANDs with the others, so one call answers "notes tagged work, updated this week, not in the Done column".New filters: board columns, custom property values, and created or updated day ranges.
Day ranges use your timezone. "Today" means today where you are, not in UTC. An MCP client, which has no browser to ask, still gets UTC.
Listings carry more. Every note in a result now shows its tags, board column, property values, dates, and whether it is a favorite, so chat stops opening a note to read one field.
Faster, longer conversations
The first answer starts sooner. A turn that needs no tools now costs one model call instead of two.
History lives on the server. Long conversations no longer grow the request, so a thread keeps working past the point where it used to fail.
Tool budgets are honest. When a question runs out of tool steps, chat is told, and it says plainly what it did not get to check instead of inventing an answer.
Stop button. Cut a long answer off mid-stream.
Tools
Both AI chat and the MCP server now expose the same 23 tools, with no exceptions on either side. A test fails the build if the two lists drift apart.
Worth knowing
get_board reads up to 20 cards per column. A busier column reports its true total and can be paged with list_notes and a column filter — chat does this on its own when you ask for the rest.