Prior now works inside your AI assistant

Our MCP server now reaches projects, notes, sharing, comments and planning, with install buttons. And an assistant can edit a note while a teammate types in it.

5 November 2026

Until now, an assistant connected to Prior could read and change your tasks and habits, write flashcards and edit budgets. Today it can reach nearly everything you can reach in the app: projects and areas, notes and folders, sharing and invitations, comments, Planning Poker, planning settings and calendar events. The server now has 76 tools, and a test fails if the number on our public page is out of date.

We also added install buttons: one-click for Claude, Cursor and VS Code, a button that copies the address and shows the steps for ChatGPT, and a command to copy for Claude Code, Codex and Gemini CLI. They are in Settings → Integrations and here:

For a walk-through of what to ask, read Manage your tasks from Claude and ChatGPT.

#What it can and cannot do

An assistant gets what a person can do in the app, with the same rules. It sees your own data and the projects shared with you, and your role in a project still applies: a viewer cannot edit through an assistant either. Plan limits apply too: starting Planning Poker, a Pro feature of the project owner, comes back as a readable error.

Some things are out of reach on purpose: your password, your sessions and access keys, billing changes and deleting your account.

We did not write a second set of rules for assistants. The tools use the paths the apps use: the same sync push for tasks and habits, the workspace sync for your own projects and notes, and the same HTTP handlers for sharing, comments and Planning Poker, run inside the server. Permissions, limits, merges and notifications stay in one place, and your open apps update at once.

Every tool is annotated as read-only, data-changing or deleting, and as reaching people outside your account or not, so clients can decide when to ask you first. The server's instructions also tell the assistant to check with you before sharing or deleting, and to treat text it finds in tasks, notes and comments as data, not as instructions.

#The interesting part: editing a note while someone types

A shared project note is a Yjs document. Yjs is a library for text that several people edit at once, and it is how teammates see each other's carets live. The difficulty is that our server is written in Go and has no Yjs client. For the apps, it only stores and relays the document as bytes.

An assistant that wants to change a shared note has two bad options: replace the whole text, which would wipe out what a teammate typed a second ago and move their caret, or stay away from shared notes. We chose a third. A small Go package, ytext, rebuilds the note's text from the stored state using Yjs's own integration rules. To change the note, it compares the old and new text and produces one update holding the smallest insertion and the deletions needed. That update is stored and sent to people in the note like a collaborator's edit. They see the new sentence appear, and their carets stay where they were, because the edit is written in Yjs's own terms and merges with whatever they typed meanwhile.

Three details keep it safe:

• The assistant sees the live text. The copy stored for lists and search can lag by a few seconds behind what people typed, so the tools rebuild the body from the document instead.
• Edits are small by design. update_note takes exact find-and-replace edits, each of which must match once, or an append, and the instructions tell the assistant to prefer them to rewriting long notes.
• Unknown content is refused, not damaged. If a document holds anything but plain strings in the note's text, the tool says so and leaves it alone.

How do we know the bytes are right? The Go tests use documents built by the real Yjs library, and the client tests apply updates written by Go with the real library.

#Limits

• The edit is written against the text the assistant read. If someone types in the same place at the same moment, what they typed is kept, and the two texts may end up side by side.
• A single edit to a shared note is limited to 256 KiB, and the tool asks the assistant to split bigger changes.
• Google and ICS calendars are read by the apps, not by the assistant. It reaches your local Prior calendars.

If an assistant does something that surprises you, tell us. The overview is in Use Prior from your AI assistant, and how it stays safe covers permissions and sign-in. Shared notes themselves are covered in Shared notes, and the tables and charts you can write in them in tables and charts in notes.