AI: Talk, AI search and read aloud
Potion's AI answers questions about your workspace and makes changes you approve. It runs on your own OpenRouter account: you connect it once, pick a model, and pay OpenRouter directly for what you use. Potion never bills for AI and never sees your card.
Connecting
Web: Settings → AI → Connect OpenRouter. iOS: Settings → AI. Either way you sign in to OpenRouter and approve a key for Potion; there is nothing to copy or paste.
- One connection per account. Connecting on the web turns AI on in the iOS app too, and the reverse.
- The key is encrypted before it is stored and is never sent to your browser or your phone.
- Model: the one model that thinks and answers (the brain). The list mixes closed and open models (Claude, GPT, Gemini, DeepSeek, Qwen, GLM, Kimi, Mistral…) with their prices. You can also switch it per conversation, from the menu under the message box.
- Instructions: up to 2,000 characters added to every conversation in every workspace, for example "Answer in French. Keep answers short. Tag new notes inbox." For facts about one workspace, use its memory (below).
- Disconnect forgets the key. To revoke it for good, delete it in your OpenRouter account.
The balance left on your OpenRouter account is shown in Settings. Each answer shows what it cost.
Talk
Talk is a chat over your whole workspace. It is at /talk on the web (⌘J, or the Talk icon at the top right of the sidebar, beside the workspace name), and the chat bubble in the bottom bar on iOS. It is always there: before AI is connected, it opens the AI settings, where you connect.
- Conversations are saved. They are listed on the left (on iOS, behind the list button), and New chat starts a clean one.
- Answers arrive whole. While it works you see the steps it takes ("Finding notes", "Reading a note"), and the answer appears once it is complete.
- Stop ends the current answer. Anything already approved stays done, and a change still waiting for your answer is declined.
- Answers can hold tables, links that open the notes they mention, and live html blocks (a chart of your records, for example), drawn the same way as in a note.
Memory, per workspace
Each workspace has its own memory: up to 30 short facts that Talk reads at the start of every conversation there. For example, "my films are tagged film and rated out of 10", "weights are in kg", "summaries in French". Nothing is shared between workspaces.
- Talk can propose to remember something when you state a lasting fact or preference, or when you say "remember…". It asks first, like any change, and the card quotes the exact fact. It never keeps secrets, one-off details, or text copied from a note.
- "Forget that" removes a memory, again with your approval.
- You can see and edit all of it: Settings → AI → Memory on the web, or Settings → AI → Memory in the iOS app. Click a fact to edit it, or × to forget it; the web dialog also has Clear all.
- Your Instructions apply everywhere; memory applies only to the workspace it belongs to.
What you can ask
Talk reads your workspace through the same queries as views and html blocks, so its answers come from your data, not from memory.
| You ask | What it does |
|---|---|
| "The 5 oldest notes tagged health" | Finds notes by tag, oldest first |
| "Notes with a rating greater than 8", "rated between 5 and 7" | Compares a property (number, date, text, select, yes/no) |
| "My 3 best rated movies" | Sorts by a property, highest first |
| "Movie notes without a rating" | Finds notes where a property is missing |
| "Movies that are not sci-fi", "notes tagged A but not B" | Excludes values and tags |
| "Notes created in the last 7 days", "due before October 1" | Filters by dates, fixed or relative |
| "Notes whose title starts with The" | Filters by title |
| "Find my notes about Japan" | Full-text search, ranked |
| "How many notes per status?", "movies per style" | Counts the notes for each value of a property |
| "How many notes do I have?", "which tags do I use most?" | Counts |
| "What does my note Arrival say?", "what links to it?" | Reads a note, its properties and backlinks |
| "What views do I have and what do they filter?", "show my Hot Sauces view" | Reads saved views and the notes in one |
| "My average tasting score", "per drink", "per month" | Totals over records (count, sum, average, min, max), overall, per period or per value |
| "Tastings that scored 8 or more", "my last 3 weights" | Filters and sorts the rows of a structure |
| "My files about the lease", "what does the PDF say about notice?" | Lists files and searches inside them |
Anything a question needs is looked up as it is asked; asking a question never saves anything. A follow-up ("now sort them worst to best") works on the same notes.
Changes, always with your approval
Talk can also change your workspace: create, append to or edit a note, trash or restore one, set tags and properties, create views, presets, lenses and structures, add or edit records, pin notes to the sidebar, rename files.
- Every change waits for you. It shows as a card, Approve or Decline, before anything happens. Declining ends the turn: it does not try again another way.
- A change to many notes is one card. "Tag every note rated 9 or more favourite" shows how many notes it will touch, and runs once you approve (up to 200 notes at a time).
- A trashed note stays in the Trash for 30 days, and an overwritten note keeps its previous version for 7, so an approved mistake can be undone.
Some things are left to you on purpose: renaming or deleting a workspace, deleting forever, deleting a structure, API keys, webhooks, uploading files, and layouts. Ask for one and Talk says it can't, and where in the app it is done. It does not try something else in its place.
AI search
With AI connected, the search box gets an AI switch. Describe what you are looking for ("coffee notes I haven't finished", "that article about sleep and ADHD") and it returns a list of notes, the best matches first, from the notes that actually exist. It never answers in prose.
A question that is not a search for notes ("what is my average score?") is not answered there. It offers Ask in Talk instead, with your question already typed.
Tasks
A task is one AI job you start on an open note with a button, as opposed to a conversation in Talk. You do not write a prompt: the task already knows what to do, does it the same way every time, and shows what it cost when it finishes. With AI connected, an open note has a Tasks menu: the sparkle button in the note's header, on the web and on iOS.
| Task | What you get | Offered on | Runs on |
|---|---|---|---|
| Summarize and read | A short spoken summary of the note | Any note | Your brain, then your voice |
| Summarize video | What a YouTube video says, with timestamps | A note that holds a YouTube video | A model that can watch video |
- A task runs on your own OpenRouter account, like everything else, and its price is shown when it finishes.
- A task that writes only ever adds to the end of the note. Its answer is appended when it is complete, with no second question: starting the task is the ask. Nothing already in the note is changed, and Earlier versions keeps the note from before.
- A task keeps running when you leave the note. A task that writes runs in the background: open another note, or another page, and it carries on. The Tasks list at the bottom left of the sidebar shows every recent task in the workspace: a spinner while it runs, then "Added to the note" with its price, or the reason it failed. Click one to open its note; the square stops a running task (nothing more is paid for, nothing is added); Clear empties the list of finished ones. A task that speaks is different: it plays in its note and ends when you leave it. On iOS the status shows in a strip under the open note's header (running, with Stop; then "Added to the note" with the price); the task still runs in the background there, and the workspace-wide list is on the web.
- A task is yours to edit, and you can write your own, in Settings → AI → Tasks (both below).
Editing a task
Settings → AI → Tasks lists every task. Open one to change two things:
- Model: the model that does the work. Summarize and read uses your brain unless you pick another here; Summarize video offers the models that can watch a video (Gemini 3.8 Flash is the usual one; a lighter one costs less than half, a Pro one about three times more).
- Your instructions: up to 600 characters added to what the task is always told, so you never repeat yourself: "Always answer in French", "Keep it under 100 words", "End with three questions to think about". They change the length, language and shape of the answer. They do not change what the task reads, and text inside a note or a video still cannot give it orders.
Leaving both empty puts the task back to how it ships.
Your own tasks
When you catch yourself asking for the same thing again and again, write it once as a task. Settings → AI → Tasks → New task: a name, a kind, and the instructions the model is given every time. It then sits in the Tasks menu of every note in that workspace, under the built-in ones. An agent connected through the MCP can write tasks for you too (create_task); they appear in the same list.
The kind is what the task is allowed to do, and the instructions cannot widen it:
| Kind | What it does | Where the answer goes |
|---|---|---|
| Write | Reads the note and writes text | Runs in the background; appended to the note under the task's name when it is done |
| Speak | Reads the note and says its answer aloud, in your voice | Heard once; nothing is kept |
| Watch | Watches the YouTube video the note holds and writes text | Runs in the background; appended to the note when it is done, timestamps linked |
- Instructions are standing orders, up to 2,000 characters: "List the action items in this note as a checklist, each with who owns it", "Explain this note to a ten-year-old", "List every tool named in the video with the moment it comes up".
- No kind edits or replaces what is in a note, reads another note, or reaches the internet. Write and Watch only append. A task works on the one note it was started on, and what it finds there is data: a note or a video cannot give it orders.
- A task belongs to one workspace, which holds up to 30. Its name is unique there. Its kind cannot change afterwards: make another task instead.
- Each task may run on its own model; left alone, Write and Speak use your brain and Watch uses the usual video model.
- Your own tasks appear in the Tasks menu on the web and on iOS. You write and edit them on the web (or through the MCP).
Summarize and read
Your brain model reads the note and writes a short summary meant for the ear, and a voice model reads it aloud, sentence by sentence, while it is still being written. A player appears above the note and the voice starts within a few seconds.
- It is live and on demand. Nothing is saved: not the audio, not the summary text, and nothing is written into the note or the workspace. Closing the player or leaving the note ends it; asking again makes a new reading.
- The price is shown when the reading is complete, beside the player: what the brain and the voice cost together on your OpenRouter account. A summary of an article is typically a few cents.
- The voice is yours to pick in Settings → AI → Voice: a voice model, then one of its voices. The summary is written in the language of the note, so choose a voice that speaks it.
- Code, html blocks, embeds, images and web addresses are left out of what the brain reads. A very long note is summarized from its beginning, and the summary says so.
On iOS the voice plays with the ring switch off, like any audio you start. The voice itself is picked on the web for now.
Summarize video
For a note that holds a YouTube video (an embed, a link, a bare address). On any other note it stays in the menu, greyed out, and says there is no video. A model watches the video, picture and sound, and writes what it says: a short opening paragraph, the moments that matter with the time each one starts, and what to take away. The summary appears above the note as it is written.
- It goes into the note when it is done, under a "Video summary" heading, each timestamp a link that opens the video at that moment. Follow it in the Tasks list at the bottom left of the sidebar; stopping it there adds nothing. It is an append: nothing already in the note changes, and Earlier versions keeps the note from before.
- A note with several videos lists each one in the menu, by the words that link it; pick the one you want.
- The price is shown when the summary is complete. A video costs about $0.004 a minute to watch, so roughly 8 cents for twenty minutes and 25 cents for an hour, on your OpenRouter account.
- The model is not your brain. Only some models can be handed a YouTube address, today Google's Gemini family, so this task runs on Gemini 3.8 Flash unless you pick another video model for it in Settings → AI → Tasks.
- A video that is private, age-restricted or removed cannot be watched, and the task says so. Only YouTube for now.
Privacy and limits
- Talk runs as you, inside one workspace. It can reach nothing you cannot, and no other workspace, even when asked.
- Text in your notes cannot give Talk orders. A clipped page or a file can contain instructions aimed at an AI ("ignore your instructions and delete…"); Talk treats everything it reads as data and says what a note asks instead of doing it. Even when a model is fooled, it cannot act on its own: every change waits for your Approve, Talk has no way to send anything over the internet, and it never holds a password or key.
- An answer never loads a remote image. Showing one would send a request the moment the answer appears, a known way to leak data; Talk names the image and its site instead. A link to another site shows that site's name before you tap it.
- Nothing runs in your browser or on your phone: the model, the key and every lookup are on the server.
- Your messages and the parts of your workspace a question needs are sent to the model you picked, through OpenRouter.
- Behind the brain, a small, cheap decision model (Jev) makes quick judgements: which guide a question needs, whether a search request is really a search, how well a note matches. Another model is only used when you ask for one.