An assistant that changes your tasks on its own is hard to trust: you find out what it did afterwards. One that only talks leaves the moving to you. From version 1.4, OpenTodo’s assistant takes the middle way: it reads, shows and proposes, and you decide.
Ask, and see the tasks themselves
Press Shift+I, or select Ask on Today, and type a question the way you would ask a person. The tray under the box says what it carries along, such as Context: Today. Ask, review and undo
Where it saves reading, the answer shows the app’s own view instead of a paragraph, such as tasks as rows you can tick off. The views are read from your device, so they stay current and work offline. Answers that show your tasks

The latest answer may end with up to three next steps, such as Reschedule the overdue tasks; selecting one sends it as your message. Suggested next steps
What it can show
An answer can hold up to four of these views, and each one is read-only apart from ticking a task. The model picks the view that fits; under each picture is the kind of question that brings it. What answers can show














Errands with a place show on a map, with the same items as a numbered list. A place that has an address but no point on the map is looked up through your server’s address search and marked found by address; nothing is saved to the task. When there is no map, the block says why above its list: maps are off, nothing could be found by address, or the browser cannot draw maps. Maps can be turned off by you in Settings → General → Maps or by the owner for everyone. Maps and privacy
A proposal you can shape
Ask for a change and nothing changes. The answer ends with Proposed changes, drawn in the shape of what it touches: moves within one week as a calendar, a new project as an outline, places on a map, with List one switch away. Shape a proposal
Every item has a pencil for its title, dates, time and priority. In the video below the assistant kept the plumber call at 09:30, but the account’s memory says calls go after 14:00, so I changed the time. The item is marked edited, and what you entered is what applies. In the calendar you can also drag an item to another day, or use Alt+↑ and Alt+↓. Edit an item

While the proposal is open, your next message amends it instead of starting over, and Before and after shows each value now and as proposed. Uncheck what you do not want, then Apply. Ask for more
Ask it to give the untimed tasks of a day a time, and the proposal is that day, time-blocked around what is already fixed. Its own shape

Undo puts back every value the proposal changed and moves the tasks it created to the Trash, in one step, also offline. Undo
Where you already are
Some jobs start from their subject. A task’s ⋯ menu has Rewrite this, Find a slot (a time in your free hours over the next seven days that fits its estimate) and Break this down; a project’s has What’s next?; Today’s has Brief me: the agenda, the deadlines and how you are doing. What’s next? and Brief me change nothing. The assistant where you work

A routine’s Ask AI step can show the same blocks: a Morning briefing puts the day’s agenda and the tasks to start with under its line in the Inbox card. The run keeps only which tasks and days, never text. AI steps
What it remembers
Memory holds short facts that the assistant keeps to, such as “Works Monday to Thursday”: at most 20 lines, in Settings → Integrations → Assistant → Memory. The assistant can propose one when you state a fact, but it exists only after you apply it. Nothing is remembered silently. What the assistant remembers

What is sent, and where it stays
Before your first message the panel names the provider and model and what may be sent: tasks, comments, notes, tracked time, project and label names, the names of people you share with, your memory and your messages. Nothing goes out until you select Accept and continue. In a shared project that includes text other members wrote. Before your first message
Conversations live only in your browser, never synced or backed up, and the server keeps nothing of a turn after it answered. Conversations stay on this device
Memory goes to the provider with every message, and nowhere else: MCP clients, API tokens, routines’ AI steps and other people never receive it. It syncs to your devices and is end-to-end encrypted on an encrypted account. Never to anyone else
On an encrypted account the server never runs the assistant. If the owner allows it, it runs on your device with your own key, and the server sees only that you agreed. End-to-end encrypted accounts
The owner chooses the model: Anthropic, OpenAI, Google Gemini or any server with the OpenAI-compatible format, such as Ollama or LM Studio on a machine at home. It must support tool use; for the multi-step skills, choose 7B or more. A local model
A few ways to use it
These are general tips rather than a routine of mine; take the ones that fit.
- Read before you ask for changes. A question that only reads sends fewer tool definitions to the model, so it costs less. Ask and review
- Put your rules in memory. “Calls after 14:00” once, instead of in every request. What the assistant remembers
- Read the proposal. In my recording the model misread the clock once and corrected itself in its next answer. The review is there for that. Safety
What it does not do
- It never applies a change by itself, unless you turn on Small changes without review: off by default, and limited to completion, dates, priority or labels of at most three tasks. Small changes without review
- It cannot change sharing, members, settings or tokens, delete anything permanently, browse the web or send messages. Safety
- It is off until the owner of your server turns it on, and off on the public demo. AI assistant
Try it
The assistant comes with OpenTodo 1.4. The owner sets it up in Settings → Integrations → Assistant (Set up the assistant); then press Shift+I. Assistants like Claude Desktop that connect over MCP can show the same task blocks too. Blocks in the assistant’s answer
