AI Assistant (Beta)
Last reviewed: 2026-11-20
Plainva can answer questions about your notes with an AI model of your choice. It reads your vault, cites the notes it used, opens notes and views for you and proposes changes — as suggestions on a note, as drafts of something new, or as a plan you confirm. It changes a note itself only where you let it (Apply directly). The assistant is experimental and switched off until you switch it on, separately on every device.
Switching it on
Open Settings → AI. While the AI is off, the page shows two cards: Plainva AI (Beta), with the switch Use AI on this device, and History and usage, so that you can delete stored conversations without switching the AI on. Without the switch there is no AI button, no AI tab and no AI window. Nothing is sent anywhere until you ask something. Switched on, the page shows its cards in this order: Models, Privacy, Without a model on this device (search by meaning, the hints next to the open note and short summaries), Other programs (on the desktop AI apps on this computer and external agents, on an iPhone or iPad Siri & Shortcuts), and last History and usage. What belongs to a single vault stands in a page of its own, Settings → AI in this vault: the card Internet, What may leave this vault, External tools, Routines, skills and memory. While the AI is off on the device, the cards stay — except Routines, which shows only once the AI is on — and a note above them says that what stands there holds once the AI is on.
Choosing a provider
Plainva does not bring its own AI service: you use a provider you choose, with your own key. Every provider can be chosen; Plainva names their terms and retention so that you decide — it excludes none.
| Kind | Providers |
|---|---|
| Cloud providers | Anthropic, OpenAI, Google Gemini |
| Gateways and own servers | OpenRouter, any OpenAI-compatible server |
| On this computer (desktop) | Ollama, LM Studio |
| On this phone | Apple (iPhone), Gemini Nano (Android) |
- In Settings → AI, in the card Models, choose Add provider and pick one. While there is no provider, the card says No model yet; on the desktop it offers exactly that button, on the phone Add provider is the row below. Each entry carries a short note on its terms — for example that Google’s free access may let people read your inputs.
- Enter the key with Enter key. The key goes into this device’s secure store; Plainva never shows it again — neither to the AI nor on screen.
- Once the first provider is connected, the choice of the model for the default profile opens by itself. Where a provider exists but that profile is still empty, the card says No model for new conversations yet and offers Choose model. On the phone the model of a profile is chosen in one sheet — provider, Search models, the list, Own model ID, for a server on this phone’s own address — not for the profiles Audio and Search model — the window and Can call tools, then Save or Clear profile. On the phone, Set up in a conversation opens the setup in three steps — Provider (On this device, Anthropic, OpenAI, Google … with your own key, or Own server), Access (the key, the phone’s own model, or the server’s address) and Try it — which ends in the conversation with Say hello. The profiles appear once a provider is there.
- Test connection (in the provider row’s “…” menu on the desktop) loads the provider’s own list of models. On the desktop a provider row shows a button only while something is missing — Enter key or Try again; replacing or deleting the key and removing the provider are in the “…” menu as well. On the phone the row itself opens these choices. If a test fails, the row says Connection failed: and why (a rejected key, no connection, an unknown model); the notice in a conversation, described below, names the sender and offers the fix.
An OpenAI-compatible server is added by its address. Plainva asks once more before it adds it, in a window of the operating system, and sends only to the address you confirmed. Plain http works only for a server on this device; everything else needs https.
If you have no key yet: a model on this computer (Ollama, LM Studio) costs nothing, and every provider’s console issues keys.
The system’s own model on the phone. On an iPhone or iPad, Add provider offers Apple (its model needs Apple Intelligence), on an Android phone Gemini Nano — on every device of that system, also where it cannot run the model; the row then says so. It needs no key and costs nothing, and nothing leaves the device — so no overview asks before sending. Its window is small, about 4,000 tokens for the notes, the question and the answer together: fewer notes go along, earlier turns are shortened, and it uses no tools. The row says whether it is ready and, if not, why — Apple Intelligence switched off, a device that cannot run it, a model the system is still preparing; on Android, Load model asks the system to download it. Apple’s model does not speak every language (no Polish). If the profile Local names it, it also writes the short summaries on the phone.
Models and profiles
In the card Models, four profiles — Fast, Balanced, Strong and Local — are your assignment of models. Choose a provider and a model for each, from the provider’s list or by typing the model id exactly as the provider names it. The profile new conversations start with carries the badge Default; another becomes the default through its “…” menu (on the phone its sheet) with Use for new conversations, and Clear profile is there too. Plainva names no model “the best”.
Under the note “More models: audio for voice notes, and the model of search by meaning.” stand two more slots. Audio holds the model that transcribes voice notes; it is never the default for a conversation. Search model holds the model search by meaning computes with when you choose Own provider under Search by meaning — see Search. The profiles and these two slots show only once a provider is there.
A model on this computer. A server on your own computer names its models, but not how much one of them can read at once, or whether it can call tools. When you choose such a model for a profile you can say both: Context window, in tokens — leave it empty if you do not know — and Can call tools. With a small window Plainva sends along fewer and shorter notes, and it stops before a request that would not fit instead of letting the server cut it silently; the notice then says how many tokens the request needs and what you stated. With the tools switched off the model answers from what goes along: without tools, without the internet and without looking anything up. A model on this device is given time: it may stay silent for up to fifteen minutes before its first word, and the conversation counts along while it works.
Fully local
Once a conversation can start, the card Models in Settings → AI starts with a line Now: … that says what holds on this device right now: that a provider answers (“Cloud”), that this device also computes something itself — search by meaning, the short summaries, reading e-mails and appointments — (“Hybrid”), or that a model on this device answers. That line is a state, not a choice: it follows from the providers and helpers you set up.
Fully local, in the card Privacy, is the one switch for it, and it is a promise: while it is on, only models on this device answer — a server on this computer, or the system’s own model on a phone —, and nothing the AI handles goes to a provider, to the internet or to another program. It applies to this device, in every vault.
- Who answers. A new conversation starts with a model of your own profiles that runs on this device: the one of the profile marked Default, otherwise the one of Local, otherwise the one of the first profile that names such a model. Plainva picks no model for you — if none runs here, the settings and the conversation say so and offer the way to set one up.
- What rests. The settings list it: providers stay set up and receive nothing, and their profiles do not answer; no conversation uses the internet; external tools and external agents are not used — a program on this computer included, because what it passes on Plainva cannot see; AI apps on this computer cannot read the vault; search by meaning through a provider stays with the words; Siri and Shortcuts are given no list of titles. Nothing is removed: switch it off again, and everything is as you set it up.
- Open conversations. A conversation that ran with a provider’s model does not go on while the switch is on — no conversation changes its model silently. It stays readable, and New conversation starts one with the model on this device.
- What “on this device” means. A server under this computer’s own address, or the system’s own model on the device. Not a server in your home network, and not Apple’s Private Cloud Compute: an address outside this device is not this device, however near it stands.
- What the switch is not. It is about the AI. Sync, e-mail, calendars and updates are functions of their own, with switches of their own. Loading a package for search by meaning is a download you start yourself; it carries nothing from your vault. And a note’s own rule — never to the cloud — holds whether the switch is on or off.
Asking
- Desktop: the AI button in the action rail, Ctrl+J (⌘J on macOS) or Ask AI in the command palette opens the conversation in the right sidebar, beside the note: it takes the sidebar’s full height, and the calendar above it shrinks to the current week. Close brings the sidebar’s sections back. Pop out turns it into a small window over your work — from then on it opens there on this device —, and Dock in the column in that window brings it back. Open as tab moves the same conversation into the AI tab. Its segments are Conversations, To decide, Skills, Memory and Agent; the history column shows only with the conversations. When something waits for a decision, the AI button in the action rail carries the number of what is to decide — exactly what that segment lists. A skill waiting for its approval, or a question of the running conversation, shows as a dot without a number, and the button’s label says that something waits for you.
- Phone: Ask AI in a note’s ⋮ menu opens the AI sheet over that note. The AI area (in the areas sheet, or in the navigation bar if you put it there) shows the conversation full screen; without an open conversation it shows a start page: To decide with its number (only while something waits), Recent (the last three, once there are conversations) with All conversations; where nothing stands there, the conversation begins at the top with Ask something. The history icon leads to the second AI screen, titled AI, with the segments conversations, to decide, skills and memory. The areas sheet shows the number of what is to decide and the bar a dot; a waiting skill or a question of the conversation shows as a dot alone, and the areas sheet says that something waits for you.
- Beside the note: on the desktop the section AI of the right sidebar opens the same conversation there, in the sidebar’s full height. On a phone or tablet it is the AI tab of the note’s context — next to Properties and Backlinks — which a tablet shows beside the note.
The note you have open goes along automatically; remove it from the context with its ✕ if you want to. Pin a note… adds further notes. The assistant can also look things up itself: it searches the vault, reads notes and their sections, databases, backlinks and linked notes, lists tasks, appointments and the notes opened or changed lately, and opens notes and views. For a look back — what changed in a period, how a note developed — it reads the version history on this device; a line you deleted it can count, but never read. Unless you let it apply changes directly (see “Apply directly” below), it cannot change, create or delete anything itself; what it can propose instead is described under “Proposing changes” below.
The assistant can also show you things: open a note at a heading, show a note in the graph, turn the calendar to a day, open views, show and hide the sidebars. It uses the commands of the command palette for that — and of those only the ones that show something: whatever creates, changes, deletes, exports or opens a window is not its to trigger.
Every conversation starts with the line “Answers are written by an AI — ⟨model⟩ via ⟨provider⟩”; below it, it says what happens to changes: “What the AI would change, it lays before you as a suggestion.” (where a folder of the vault has “Change without asking”, it names that exception) — or, while Apply directly is on, that the AI may change and create notes here. The button at the foot of the conversation names the profile (for example Model: Balanced) and opens the profile menu. Before the first message, the row above the input carries labelled chips, the switches first: Internet (only where the vault allows the internet), Apply directly (see below) and My style (only with a writing style), then Pin a note… and What goes along? (the view of the context, see below). A switch that is on is filled and carries a check; on the phone the row is one line you slide sideways. After the first message internet and style can no longer be switched — Apply directly can —; the marking line says what holds. In an empty conversation Skills under the input field lists the skills, the featured ones first; on the phone its last entry, Skills and memory, leads to the skills and the memory. An empty conversation without a start page begins at the top with Ask something. The input says “Ask about “⟨note⟩” or your vault…” with a note open, otherwise “Ask about your vault…”; the send button sits in the field. On the phone the foot gives way while the keyboard is open.
Under each answer, Copy copies the answer and Keep as a note keeps it (see below); under the last answer stand Learn from this conversation and a “…” menu with Same question to another model…, which leads to Details. Beside them one short line says what was sent where: to which provider, how many notes and — where the provider publishes prices — the approximate cost. Details opens the full overview of the run: tokens, cache, memory, style and pages. Stop ends an answer at any time.
A question that waits for you — the send overview, a page, a search, a tool, a question from a server — stands above the input field, not in the scrolling conversation; the field then says “Answer the question above first”. However long the question, its buttons stay in view at its foot — in the sidebar too.
If something fails, the message names who it comes from — this device, the network, the provider or Plainva — and offers the exact fix: Enter a new key for a key, Change model for an unknown model. A provider’s own error text is behind Details. Offline, the notice says that this device is offline and to send the question again.
A link in an answer opens only after you confirmed its address, and images in answers are never loaded.
What goes along
With every question Plainva puts together what may matter — on this device, before anything is sent:
- Where you are: the date and time, the note you have open (on the desktop also a database) and your selection in it — for a note also its properties, the first sixteen, empty ones left out —, open tasks due in the coming week or overdue, the appointments from two hours ago to the end of tomorrow, today’s daily note and, on the desktop, your open tabs.
- Notes that may matter: found through your words, the links of the open note and what you opened or changed lately. Your privacy rules decide first; only the notes they allow are ranked at all. A few go as sections — not as whole notes —, others only with their title and a card — the first sentence of their section and every sentence with numbers, dates, tasks, negations or links, word for word — or just their name; the assistant reads more of them when it needs to.
A section or a whole note that this conversation already carries and that has not changed since is named, not sent again; cards and the situation above go anew with every message. Place stamps from your journal and mood values are never sent on their own.
Before anything is sent
The first request of a session shows an overview, titled “Send to ⟨provider⟩?”: where it goes, which notes and which part of each (with leave out and redact), what else goes along (your selection, appointments, tasks), the internet and the size. The memory and the style that go along and what was kept back stand there too; the instructions and the tools are behind Details. Send sends it; Cancel sends nothing and gives your words back to the input; leaving the conversation while the overview is open — opening another one, starting a new one or deleting this one — does the same; the − next to a note leaves it out. Within what you approved, the next requests go without asking. The overview comes back whenever the scope grows: another model or provider, a new kind of data, notes from another folder, new tools, the internet where it was off before, a kind of sensitive text that no approval covered, instructions of your vault that were not part of what you approved, or a much larger request. A model on this device never asks.
If you want to see the overview before every request, switch on Ask before every request — in its own line above the buttons of the overview, or in Settings → AI in the card Privacy. On the phone the buttons of the overview are stacked.
Under each answer one quiet line says what it is based on — the notes that went with it, by name — and to whom it went; Details opens the overview of what went with it, with the answer’s coverage. A warning appears only when an answer names a note as its source — a link to it — that was neither sent nor read: check that statement against your notes. If the vault could not say in time where an answer’s links lead, Details says that this was not checked.
What made an answer. Opened, Details also keeps what the answer rests on: when you asked, each note with the version that went — a short code — and whether it is unchanged, changed since or no longer there, the notes the assistant’s tools read, the tools it used, the version of the skill it ran with, and what it left for you to decide. Plainva keeps paths and versions, never the text; whether a note changed, this device checks when you open the details. Ask another model… (in the opened Details; the “…” menu leads there) asks the same question with the same notes in a new conversation: nothing a search would find is added, the note you have open now stays out, and so does what your privacy rules keep from that model now. The send overview shows what goes, as for every request, and the new conversation names the one it comes from. This works for the first question of a conversation — a later one rests on what was said before — and for questions that brought nothing but notes along.
What goes along?
The chip What goes along? above the input shows what the next request would carry — before it goes, for the model chosen now. For every note: why it was chosen (open now, pinned, matches your words, close in meaning, linked, due soon …), which part goes and roughly how many tokens. For each note you can
- leave it out of the next request (Take back brings it back),
- pin it to the conversation,
- keep it on this device for good: that writes the rule
cloud: denyinto the note (see below).
Notes your rules keep back are listed as well, so that you know what is missing; they are never scored and never sent. Send with this context sends what you typed. In a wide AI tab the view stays open as a column beside the conversation. What you pin or redact here while the send overview is open or an answer is still on its way counts from the next message on, and stays with the conversation.
Above the notes, Sent says how much of these notes goes along — for example ~870 of 3,460 tokens — and Saved how much less that is than sending every proposed note whole; the first time, it also says how many tokens that would have been. Show as a trail in the graph opens the graph with the open note and the sources ringed and the links between them.
When the text that would go to a cloud looks like it holds a password or key, an account or card number, an ID or tax number, or health details, a line under the note says Possibly sensitive and what it saw — in What goes along? and in the overview before sending. In What goes along? the line also offers Keep on this device, which writes the rule cloud: deny into the note; the overview offers only the redacting. For numbers and secrets, Redact in this conversation replaces them with a placeholder such as ⟦withheld account⟧ in every message of this conversation, also where the model reads the note itself, until you choose Send unredacted; the overview counts them under Kept back. Tasks and appointments have the same choice in the line Tasks, appointments and the open note’s details. The first time in a session that something of such a kind would go unredacted, the overview asks before sending. The check runs on this device and is a hint, not a filter: it can miss things, and it never stops a request. A passage sent by a suggestion action goes as it is, and the overview only names what it holds; a selection that goes along with a question is checked like a note and can be redacted too. With a model on this device no hint appears.
Short summaries
With Short summaries with the local model (in Settings → AI, card Without a model on this device, off until you switch it on), a model on your computer writes short summaries of the longer sections of your notes, of whole notes, of the top-level folders and of the vault. It runs only when the profile Local names a server on this computer (Ollama, LM Studio; on the phone, the system’s own model) — never a cloud in the background — and only while Plainva is idle; on the phone only while it is open. Every short summary is checked: a section’s summary must keep every number, date, amount, link, tag and negation word for word, otherwise the section’s own sentences go instead. A short summary is bound to the exact text it stands for; after you change the section it is not used until it has been written again. Folder and vault summaries are written only from notes your rules let go to a cloud. In What goes along?, a source sent as a short summary says so, and Original sends its own sentences with the next message. To a model in the cloud and to a conversation that may use the internet, a short summary goes only where its note holds nothing back from that reader — no withheld link, in its text or its properties, nothing else kept back —, and a folder’s or the vault’s only where none of its notes does. That is decided when the message goes: a summary written before a link was added or a note was kept back does not go.
With a selection
Select text in a note, and the AI works on just that passage.
- Desktop: while you edit, AI in the selection toolbar offers As a suggestion — Rewrite, Shorten, Translate…, Tasks from it — and Ask in the conversation — Explain and Ask about the selection… (Ctrl+J, ⌘J on macOS).
- Phone: AI in the bar above a selection — when reading as when editing — opens the AI sheet.
- In every conversation: as long as text is selected in the open note, the row With the selection above the input offers the same actions, except Ask about the selection… — that is what you type in the field.
A suggestion action sends the selected passage with the name and path of its note — not the rest of the note, no pinned notes, no memory, no tools — and asks with the same overview as a question. The answer comes back into the note as a round of suggestions, like one from a person: under Suggestions you accept or decline each change or the whole round, and nothing in the note changes before you do. The round’s author line reads Plainva AI · ⟨model⟩, so it stays visible which passage an AI wrote. Tasks from it adds the tasks below the passage instead of replacing it. Each action keeps its conversation in the history.
A passage from a note your rules keep from the cloud — or one with links to such notes or with place stamps — goes to no cloud model. In an encrypted workspace the suggestion actions are not available yet: its suggestions cannot name the AI as their author yet.
In a comment thread
Address the assistant in a comment and it answers in the thread. Type an @ in the comment field and pick AI — the entry with the AI mark — or type the name yourself: @AI, @KI and @IA all reach it, whatever language the app speaks. Once your comment is sent, the thread shows the row is writing a reply… under AI; Stop ends it. The answer appears as a reply in the same thread, authored Plainva AI · ⟨model⟩. Unlike a suggestion it does not wait to be accepted — it is a remark beside the note, never text in it — and on the device that asked you delete it like one of your own.
The thread goes to the model the way a question does: its comments, the passage it hangs on and the note itself, through the same overview. A comment thread is a kind of data of its own, so the overview asks the first time. Where your rules keep the note from the cloud, its comments do not go there either, and links in them to such notes are withheld. Only a comment you send on this device calls the assistant; one that arrives through sync never does, whatever it says. Web addresses the AI brings along itself — in a reply, a suggestion or a transcript — are written so that nothing opens or loads them (https[://]…); addresses your own text already contained stay as they are. In an encrypted workspace the assistant cannot be addressed yet: its comments cannot name the AI as their author yet.
Skills
Skills are instructions for recurring work. Fourteen come with Plainva — among them Daily orientation, Weekly review and Project status, which Skills under the input field of an empty conversation lists first — and you can write or import your own. Start one with a click, or simply ask: the AI loads a matching skill by itself. Your own skills run only after you approved them on this device. Everything about them: Skills.
Scripts are small programs you write for what a model does badly — counting, sorting, adding up. They run in a closed box, read your vault through the same tools as the AI, and only after you approved them on this device. Everything about them: Scripts.
Routines
A routine starts a skill at a time you set — the morning brief on working days at seven, the weekly review on Friday afternoon. You set one up in the vault’s settings under Settings → AI in this vault, in the card Routines: New routine, then a name, the Skill, the sentence it is told each time (left empty: the skill’s own request), the days and the time, what happens If Plainva was closed at that time — Catch up once or Skip —, and the Model. Only skills that are active on this device and only read can become a routine. If the vault has a folder with Change without asking, so can skills that change or create notes.
The form’s button On to the overview leads to a confirmation titled “How “⟨name⟩” runs without you”. Before a routine exists, Plainva shows what holds for a run nobody watches, and only your yes — Set up routine — makes it. A routine only reads — no suggestion, no draft, no internet, no external tool, no script, no command of the app —, except in folders with Change without asking (see Apply directly): there it may change and create notes, never delete, rename or move them, and the confirmation says so. What a run changed stands in the routine’s row under Changed:, with Undo. It has a small budget. Its result is a note in the vault’s Inbox folder — a folder called Inbox until you name another in Settings → Notes & filing —, called after the routine and the day, marked as written by the AI, with the notes it rests on. With a model on this device nothing leaves it. With a provider’s model you confirm what a run brings along by itself — the overview shows what that would be today —, and a run that would bring along more than that sends nothing and waits for you. What the model then reads through its tools is decided by your privacy rules, as in every conversation. A routine does not read e-mails for a provider’s model: Plainva asks about that anew in every session, and nobody is there to answer.
A routine never asks a question. Where it would need an answer — more would go than you confirmed, this device asks before every request, its skill was changed since you confirmed the routine, the provider of its model is no longer set up, or its model is not on this device while Fully local is on — its row says what it needs, with Look at it. There you see what a run would send now and what of it is new, and Confirm again lets it run again.
Routines run on this device while Plainva is open: shortly after you open the vault, then about once a minute, and when the app comes back to the front. Plainva has no service in the background and promises no exact time. A run that was missed while Plainva was closed is caught up once at the next opening — however many were missed — or, as you set it, skipped: that holds for the moments of an earlier day, a moment missed today still follows. With Only when something changed the device first checks whether a note changed since the routine last left a note, and asks no model otherwise. A routine belongs to the device you set it up on: it is kept in the app’s data, not in the vault, so no other device runs it.
While a routine runs, a line in the conversation says so, with Stop, and a message of your own waits until it is done. The switch in its row pauses a routine; its menu runs it at once (Run now), edits it or removes it. The notes it wrote stay. Each run leaves its conversation in the history under the routine’s name, and the next run replaces it — unless it is open on screen, or you went on with it: a message, a new title, a note you pinned or redacted, another model. Then it stays.
A routine can also keep a folder’s log. Choose Keeps a folder’s log in the form and the Folder — or The whole vault. At its time Plainva looks up on this device which notes in that folder, or below it, are new or were edited since the last entry, and shows a model, of each of them, the lines it gained or its beginning — at most twenty notes per run, in one request, without any tool. The model writes one sentence per note. The entry itself is written by Plainva: under the day’s date at the top of the folder’s log.md, one line per note with New or Updated, the link to the note and the model’s sentence. That is the log.md of the Open Knowledge Format; Plainva creates the file where it is missing and leaves everything else in it as it was. When nothing changed, no model is asked and nothing is written.
A log says nothing that a note’s own rules keep back. log.md has no rules of its own; the rules of its folder hold for it. A note that carries stricter rules than that is left out of the log entirely — it is not named, and no model is shown it for the log. What was removed from a note is never shown to a model and never written down: a line you deleted stays deleted. A folder your rules keep from a provider needs a model on this device for its log. Keep a folder’s log on one device only — otherwise two devices write the same entry.
Hints
Beside the open note, Plainva shows up to three things it notices by itself — on the desktop in the section Hints of the right sidebar, after Related; on the phone in the tab Hints of the note’s sheet. A link of the note that leads nowhere: Create note creates the note where the link will find it, as a click on the link would. A line that begins with TODO and is no task yet: As a task turns exactly this line into a task where it stands (- [ ] …) and changes nothing else in the note. And in today’s daily note, an appointment of tomorrow that no note is connected to: Create meeting note does what the calendar does. Nothing happens without your click, no model is asked, and nothing leaves this device. A mention of another note without a link is offered by the context graph, related notes by Related — neither appears here a second time.
Why this hint? says the rule in one sentence and leads to the line; Don’t show again hides exactly this hint on this device — if the line changes, it can come back. Pause for this note and Pause in this vault are in the section’s menu, on the phone under the list. Show hints at the open note under Settings → AI, in the card Without a model on this device (together with Show related notes), switches the hints off on this device; there you also resume a paused vault or paused notes and bring back hidden hints.
Memory
What the AI should know about you without being told again — what you do, how you like your answers — is kept in the vault’s memory: two files you can read and edit. What stands under Always included (the file names are behind the ⓘ of that card) goes into every conversation you begin — not into Explain image, Make a note of it, the reply in a comment thread, the actions on a selection, a skill’s test run or a folder’s log, each of which answers one question where it was asked —, the rest is looked up when a question needs it, and nothing gets into it without your yes. See Memory. There, too, is your writing style: My style under the message field gives one conversation what this device measured in notes you chose.
Learning from a conversation
A conversation can teach the AI something for later. Learn from this conversation — under its last answer — sends it once more to the model that led it and brings back drafts: entries for the memory, rules, a new skill, or other instructions for one of your skills. What goes is your words, the answers, the names of the tools and how many pictures a message carried — no tool result and no note that went along as context. Where the conversation used a skill of your own — one you wrote in this vault, not an imported one and not one that comes with Plainva —, its name, its description and its instructions as they are now go along too, because a suggestion may give it other instructions. A long conversation is shortened: each message at 4,000 characters, the whole at 60,000, the middle left out. A conversation that read text written by strangers (a web page, a mail, a server) or that rests on a note kept from the cloud or from the internet proposes memory entries only and sends no skill; with a provider’s model, a conversation that rests on a note kept from the cloud is not sent at all. Nothing of it counts before you accept it. See Skills.
Transcribing a voice note
At every voice note — in the editor, in reading mode, in the journal and on cards — Transcribe turns the recording into text. It goes as it is — with its file name where the provider takes the recording as a file (OpenAI and compatible servers), without it where the model takes it inside the request (Gemini) — to the model of the profile Audio, through the same overview as a question; a recording is a kind of data of its own, so the overview asks the first time. The transcript comes back as a suggestion under the recording, authored Plainva AI · ⟨model⟩ — accept or decline it under Suggestions. With Summarise transcripts on this device (in Settings → AI, in the card Without a model on this device), a long transcript gets a short summary above it — written by the model of the profile Local on this device and checked: it names no number, date or link the recording does not. Without such a model, or when the check fails, the transcript comes alone.
Audio needs a provider with an audio route: OpenAI (for example gpt-4o-transcribe or whisper-1), Gemini, or a compatible server of your own — one on this computer keeps the recording on the device. Recordings up to 11 MB can be transcribed. A recording in a note your rules keep from the cloud goes to no cloud model, and encrypted workspaces do not offer it yet.
Explaining an image
At every picture of the vault, Explain image asks the AI what it shows. Make a note of it, beside it, asks for a note instead: a heading, what the picture shows, the text in it word for word and, where it shows things to be done, a task list. The answer is the note — it comes into your vault when you press Keep as a note.
- Desktop: in the toolbar of an open image, and in the menu a right-click on a picture in a note opens — while editing and in reading mode.
- Phone: below an open image; at a picture in a note, Open image takes you there — or press and hold Open image for a menu with Explain image and Make a note of it.
The picture goes, with the question, its file path in the vault and — where you started from a note — that note’s title, to the model new conversations start with — in a conversation of its own, where you can ask on: what a table says, what stands in the second column, what a diagram means. Like every question, it also brings what goes along with a question (see “What goes along” above): the note you have open, the notes Plainva finds for the question and where you are — open tasks, appointments. The overview shows the picture and these before anything is sent; a picture is a kind of data of its own, so the overview asks the first time.
What goes is not the file. Plainva draws the picture, scales it down to at most 1,568 pixels on its longer side and saves it anew for sending. So it goes without what the file records about it: the place a photo was taken, the date, the camera. The overview shows exactly the picture that goes, with its size. That copy stays with the conversation on this device, so you can still see later what the provider got; delete the conversation and it is gone.
Rules. A picture in a folder your rules keep from the cloud goes to no cloud model. Neither does a picture that a note with the rule cloud: deny shows — wherever you press Explain image, also at the open image: before sending, Plainva looks up which notes embed the picture, and if it cannot find that out, the picture stays. A model on this device stays allowed. What is written in a picture is content, like a note’s text, never an instruction: the conversation of Explain image can look things up in your vault, but it cannot use the internet and triggers nothing in the app.
Which models read pictures. Most cloud models do. The system’s own model on the phone does not, and Explain image says so. Where a provider’s list says that a model reads no images, the overview tells you before you send. If a provider turns the request down, choose another model below the conversation and ask again — the picture is still in it.
On the internet
The assistant cannot use the internet until you allow it — three times over:
- For the vault. In Settings → AI in this vault, in the card Internet, switch on The AI may use the internet in this vault. It is off for every vault until you decide, and it applies on this device only.
- For a conversation. Before the first message of a new conversation, press the chip Internet above the input field — Let this conversation use the internet. Whether a conversation may use the internet is decided when it starts; to change it, start a new conversation. A conversation that may use it says so in its first line. Starting the skill Research is the same choice: its conversation may use the internet — see Skills.
- For every request. While your notes are in the conversation, every page the assistant wants to read and every search it wants to make asks first, with the whole address or the search words — that is everything that leaves your device for it. Read page or Search lets this one request through; Don’t allow leaves it, and the assistant goes on without it.
What a request is. Reading a page is one request from this device to the site, like opening the page in a browser — without cookies, without a login and with nothing from your notes; as with any visit, the site sees your internet address and that the request comes from Plainva. Only public pages over https are read; addresses in your home or company network are refused. A search goes to the provider of your model — Anthropic, OpenAI, Google Gemini or OpenRouter — which searches with exactly the words you were shown; providers may charge for searches separately. A model on this device can read pages but cannot search, and the system’s own model on the phone cannot use the internet at all.
Where an address comes from. The question says whether you named the address, whether a note or a result named it — or whether the model put it together itself. An address the model composed could carry something from your notes in it: read it before you let it through.
Sites that need no asking. With Always for ⟨site⟩ in a question, or under Sites that need no asking in the vault’s settings, pages of a site are read without asking — as long as you, a note or a result named the address. An address the model composed always asks.
What the assistant reads. Never the page itself. A second request to the same model, one that has no tools, reads the page and writes a short report: a summary, statements with the passage they rest on, and links that really are on the page. A page that tries to give the assistant instructions therefore reaches it as a report about a page — never as a page it works on. Under the answer, Read on the web lists the pages that were read, and the line under it opens everything that was asked for.
Notes that stay out. A note or folder with not allowed under Onto the internet (see What may leave this vault below) does not exist for a conversation that may use the internet: not in its context, not for its tools, and links to it are withheld.
A link in an answer whose address the model composed itself is marked, and the question before it opens says so. If no answer comes back at all — no connection, the provider does not respond — the conversation lists the notes that match your question best instead.
Mail and appointments
Appointments. The assistant lists appointments from your connected calendars — day, time and title, on request also the place and who takes part — and reads a single appointment in detail: the organiser, the attendees with their answers, and your own. The link that a calendar attaches to an online meeting is never handed over; the assistant only learns that the appointment is online. A link that stands in the place or in the description is part of that text: the place goes along as it is, the description goes to the reader described below, who may name links that are really in it.
Mail. Where mail accounts are connected in this vault, the assistant can search and read messages. Mail is not one of the tools a conversation starts with: the assistant looks for it only when your question needs it, and at the first access Plainva asks — Read your mail? Allow holds for this provider until you close Plainva; another model or another provider asks again. Don’t allow leaves the access out, and the assistant carries on without it. A model on this device does not ask, because nothing leaves the device for it.
What the assistant reads of it. Of a search it sees the date, the sender and the subject of the messages, and whether they are unread, flagged or carry attachments — never their text. Of a message it opens it sees the same, the To line, the account and the folder it lies in, and the names, types and sizes of the attachments, but not the text. When a search covers several accounts, every line of it names its account; a line that says an account could not be reached names it too, and where a folder name fits no folder of an account, the line lists that account’s folders. It never reads the text of a message or the description of an appointment itself: other people wrote them, and whoever writes a mail or an invitation can write it for exactly this reader. A second reader without any tools reads them and writes a short report — a summary, statements with the passage they rest on, and links that are really in there. If the conversation itself runs on a model on this device, or a model on this device is set up as the profile Local in the card Models, it is that reader, and the text itself does not leave the device; only the report goes to the provider. Otherwise the conversation’s provider reads it, in a request of its own without tools. For mail, the question tells you beforehand who reads; the description of an appointment goes to its reader without a question, and the line under the answer says who read it.
What does not change. The assistant only reads: a message it read stays unread, nothing is moved, answered or deleted, and it does not open attachments — it only names them. Below the answer you see how many messages were read, and the line below it says who read the text. An e-mail or an appointment it writes for you is only ever a draft that you send or save yourself — see Proposing changes below.
External tools
The assistant can use tools of servers you connect yourself, over the Model Context Protocol (MCP) — a ticket system, a wiki, a database of your team. It is the other direction of Connecting AI apps: there, other apps read your vault through Plainva; here, Plainva’s assistant asks other servers. Nothing of a server is used before you have looked at what it offers, and every call is shown to you before it goes out.
Adding a server. In Settings → AI in this vault, in the card External tools, choose Add a server…. Give it a name of your own and its address (https://…; http:// only for a server on this device), and an access token if the server asks for one — it goes into this device’s secure store and is never shown again. On the desktop a server can also be a Program on this computer: the file to start, its arguments and the values for its environment. Plainva starts it directly, without a shell, and in a sandbox where your computer has one Plainva can use — Run it in a sandbox in the dialog switches that off. Your system shows the address or the whole command once more before it is remembered. On the phone a server is always an address.
Signing in. Some servers want a sign-in instead of a token. Its review then says The server asks for a sign-in. Choose Sign in…: Plainva asks the server where its sign-in lives, opens that page in your browser — on the iPhone in the system’s sign-in window — and waits until you are back. What it receives stays in this device’s secure store and goes only to this server; neither you nor the AI ever sees it. It is renewed without you for as long as the server allows, and when it has ended, the review asks you to sign in again. Where the sign-in service does not let apps register themselves, Plainva asks for the Client ID the operator of the server gave you. Sign out forgets the sign-in; a sign-in and a stored access token replace each other.
Reviewing it. A server that was just added offers nothing yet. Its review shows what is registered and what the server lists: its own description, its tools with their descriptions — the server’s own words — and its Text templates of the server. Approve allows exactly these texts, on this device. Before a server is used — unless the list it gave is still fresh by the server’s own account — Plainva loads what it lists again and compares it with what you approved; if anything differs, the server is blocked until you look again, and the review says what changed.
What a vault allows. Each vault decides for itself: whether it uses the server (Use ⟨server⟩ in this vault), which of its tools the assistant may call — none is ticked —, and under Notes that may go with a call whether None, Chosen folders or The whole vault. A tool that does not say it only reads can be ticked as well; its row says what a call can then do at the server — change something, or change, overwrite or delete something. A tick holds for what the tool said when you set it: if the tool later says it can do more, it is not offered again until you tick it anew.
In a conversation. The tools of your servers are not among the tools a conversation starts with: the assistant looks for them only when your question needs them, and the overview before sending names the servers they belong to. Every single call asks first — Call ⟨server⟩? — with the tool and exactly what would be sent. Call lets this one call through, Don’t allow leaves it, and there is no “always”. A tool that can change something asks in other words — Let ⟨server⟩ change something? —, says that Plainva cannot undo it, and its button reads Run it. A call does not go out at all if the conversation has read a note that lies outside what the vault allows this server, or one you keep from the cloud. What comes back is treated as a stranger’s text: the assistant reads it and takes no instructions from it.
Questions from a server. Some servers need a detail from you during a call — which project to search, say — or want you to sign in at a service they reach for you. In the server’s review, under In this vault, Let ⟨server⟩ ask you questions allows that; it is off until you tick it, and a server that asks anyway gets no answer. A question stands above the input as Plainva’s own card — ⟨server⟩ has a question —, with what the server wrote as its quote and the fields it asks for: Answer sends what you filled in — where it asks for nothing, Confirm says yes —, Decline says no, Cancel stops. A server that wants you to open a page gets the card ⟨server⟩ would like you to open a page, with the whole address and its domain apart; Plainva opens it in your browser only when you choose Open in browser, and the call goes on with Continue. Plainva gives your answer to that server alone and never to the AI — what the server sends back may still mention it; the history only notes that a question was asked and how it ended. A server may not ask for passwords, keys or payment data this way: Plainva does not show such a question — Plainva blocked a question from ⟨server⟩ says what was asked —, and the server gets a no.
Text templates. A server can offer text templates (its “prompts”) — ready-made requests. They are in the menu Skills of an empty conversation, and only you start them. The first time, Plainva shows what a template expands to before it is sent as your message; from then on exactly that text goes without asking, and another text blocks the server.
Skills. A server can also bring skills — guides for its tools. Each one waits under Skills, From servers, until you approve it there; see Skills.
What Plainva keeps. The address or the command is remembered on this device, the stored values in its secure store; your approval lies in Plainva’s own data, never in the vault — so whoever can write the vault cannot approve a server. Under Recent calls in this vault the review lists when a tool was called, which one and how it ended — never what was said. Remove server deletes the server from this device, for every vault.
A conversation a skill started, an action on a selection and a reply in a comment thread do not reach external tools, and neither does the system’s own model on the phone.
External agents
On the desktop, Plainva can also start an AI agent of another maker in the vault’s folder — a program you installed and signed in to yourself. Such an agent is not the assistant: it reads and sends on its own, and your privacy rules and the overview before sending do not reach what it does itself; only a file it asks Plainva for is checked against your rules. What Plainva controls in its session and what it does not: External agents.
Proposing changes
The assistant can propose more than an answer — and, unless you let it apply changes directly (see “Apply directly” below), nothing it proposes is in your vault until you say so. There are three forms, and each waits where you decide about it.
- A suggestion on a note. Ask for a change to a note that is there, and the assistant lays it on the note as suggestions: in the margin, signed “Plainva AI · ⟨model⟩”, each change to accept or decline on its own — like a person’s suggestions, see Comments & Suggestions. Accepting keeps the note as it was as a version first, so the version history always has the way back. Under the answer a line names the note; press it to open the note. A value for one of the note’s properties is proposed the same way: the card shows the property with what it says now struck through and what it would say, and marks a property the note does not have yet as new. If the property says something else by the time you decide, the card tells you that the suggestion no longer fits. Who made a note and who vouches for it is nothing the assistant can propose (see OKF), and neither are the properties Plainva keeps for itself. In a database, a value proposed for an entry also stands in the cell of that entry, where you accept or decline it without opening the note — see Databases (.base).
- A draft. A new note, a task or a journal entry is left as a draft: a card under the answer says what it would become and where it would go. Create makes it — the note in the folder the card names (the Inbox folder, unless the assistant named another), the task read from its words as if you had typed them into the capture field, the entry in the journal of the day on the card. Show unfolds a note’s text first; Discard throws the draft away. A note made from a draft says who wrote it (
generated, see OKF) and names the notes the conversation rested on. Where your new tasks also go to a task list of your provider, a task’s card carries the capture field’s switch, Also create in “…”: it is on, and the task is created there as well unless you switch it off. An entry of a database is drafted the same way: its card names the database and the properties the entry would have, and Create writes it as a note into the folder that database keeps its entries in — with those properties and whatever makes a note an entry there. A database that has no storage folder for new entries yet takes such a draft only once you have created its first entry yourself. - A plan. Renaming, moving or deleting a note cannot be reviewed piece by piece, so the assistant asks: a question above the input field shows what would happen — the new name and how many links in how many notes follow it, or the target folder, with a warning if the note would leave a privacy rule of its folder behind. After your yes, Plainva does it the way it does when you do it yourself; the assistant only learns whether it happened. For a deletion the question merely opens Plainva’s own delete dialog: nothing is gone before you confirm there. One of a note’s own privacy rules is asked about in the same way and is never laid down as a suggestion: the question names the rule and whether it would be written into the note or taken out — with a warning where taking it out lets the note go to cloud models or into conversations with the internet again. After your yes Plainva writes it; the rule holds from then on and does not take back what a conversation has already sent.
An e-mail and an appointment. Where a mail account or a calendar that takes appointments is connected in this vault, the assistant can also draft an e-mail or an appointment. It sends and saves neither. The card names everyone it would go to — To, Cc and Bcc, or the Attendees — and points out every address you did not write yourself in this conversation: the assistant may have it from a note, a mail or a web page, so check it. Open in Mail opens the mail as a new message with everything filled in, Open in Calendar opens the appointment in the calendar’s event editor; there you change what you like, and Send or Save is your own step. Attendees get an invitation from your calendar provider when you save, as with every appointment you enter yourself. The draft stays in the list until the mail has really gone out or the calendar has taken the appointment: closing the message, taking a send back within its few seconds or a calendar that refuses leave it where it was. A mail you store with Save draft lies with your mailbox’s drafts from then on and leaves the list as well; on the desktop so does a message you move into a window of its own.
At a database the assistant also works without a conversation: Fill “…” with AI… reads the notes of the entries that have no value in a column — up to 25 entries per run, at most 6,000 characters of each note — and suggests one for each, and in the filter settings a sentence becomes filter rules that you see before they apply. Before anything is sent, the overview shows what goes, as always — for a column the notes of those entries, each in a request of its own; for a filter only the database’s columns with their names, kinds and choices, and your sentence, never an entry. Both are described under Databases (.base).
Everything that waits is in one list: To decide, a segment of the AI tab on the desktop and of the AI screen on the phone. It names the notes that carry suggestions of an AI and the drafts of this device, each with who left it. Drafts are kept on the device they were made on, like conversations; suggestions belong to the note’s comments and reach your other devices with them.
Three limits hold whatever the assistant is asked. A web address it brings into a suggestion or a draft is written so that nothing opens or loads it (https[://]…); an address you typed yourself stays as it is. An HTML attribute that would load one gets its name in brackets ([src]=…). Where the assistant changes a note, only what it changes is checked: your own code and HTML elsewhere stay as they are. Only a <style> or <form> before a change, and a <base> anywhere in the note, is written as text — what follows would otherwise be read as a style sheet, as part of a form or as an address against that base. A conversation that has read a note kept from the cloud or from the internet lays a suggestion, a task or a journal entry only where the same rule holds — a drafted note or database entry takes the rule along instead, and an e-mail or an appointment is not drafted at all. And in an encrypted workspace nothing is proposed, drafted or planned.
The assistant is asked to name a note as a link, so a link is the one claim in its text that Plainva can check: where a suggestion or a draft links to a note your vault does not have, the line under the answer and the draft’s card say so — Linked, but not in this vault: and the names. Nothing is held back for it; you may want the link first and the note later. A link to a note that is in your vault but that the AI may not read here — a privacy rule keeps it from the conversation — gets a line of its own: Linked to notes the AI may not read here: and the names. The AI itself is told the same about both kinds, so it cannot find out a name by trying it. Whether a statement is true, Plainva does not check.
A skill has these abilities only when it names them, and its review says so — see Skills.
Apply directly
Out of the box the assistant changes no note itself — everything above is a suggestion, a draft or a question. For routine work (“clean up the links in these ten notes”) you can let it apply changes directly, in two ways:
- For one conversation: the chip Apply directly in the footer of the conversation, beside Internet and My style, switches it on for this one conversation — before the first message or in the middle of it. While it is on, the line above the conversation says: The AI, and scripts it starts, may change and create notes in this conversation. Deleting, renaming and moving still ask. After Plainva restarts it is off again for every conversation; the overview before sending names it under Apply directly.
- For a folder: under Settings → AI in this vault, in What may leave this vault, every folder rule has the column Change without asking (on the phone a highlighted line under the folder, only where it holds there — you change it by tapping the row). Set to allowed, the assistant applies changes there directly, without the chip; on a note from that folder the chip then reads Apply directly · folder. not allowed takes it back for a subfolder. The rule is kept in
.agent/policy.ymland travels with the vault like the other rules.
Applied directly are a change to a note and to its properties, a new note, a database entry, a task and a journal entry; this also holds for what a script writes that the assistant started in the conversation (see Scripts), while a script you start yourself with Run never writes directly. Renaming, moving and deleting still ask, and so do a note’s own privacy rules; an e-mail and an appointment stay drafts. What an external agent hands to Plainva, what an AI app proposes through the MCP server and the actions on a selection always stay suggestions or drafts. What an agent writes itself, as a program of its own, is in the vault at once (see “External agents” above).
Every direct change has a net. Before it, Plainva keeps the note as a version, whatever the interval between backups; where that fails, the change becomes a suggestion. What the assistant wrote stands in the note’s version history marked AI. Under the answer, Changed: says how many notes — each with View and Undo, plus Undo all. Undo puts back the version from before, and takes away a note the assistant created. If the note has changed since, Plainva overwrites nothing: a notice offers Compare and opens the version history.
A change still becomes a suggestion when the note is open in the editor and not saved yet, when it has an open sync conflict, when it changed while the assistant worked on it, or when it may not be written here. The line under the answer then says why — for example “Planning” was open in the editor with unsaved changes — the change waits there as a suggestion. If the note is open and saved, the editor takes the change at once.
Keeping an answer as a note
Under every finished answer, Keep as a note turns the answer into a note of your vault. You press it and Plainva writes the note — the assistant itself still changes nothing. A note kept from an answer about a picture shows that picture, and an answer that opens with a heading of its own gives the note its name — unless the name is set already: the note kept from the first answer of a conversation that a skill started is called after the skill and the note that was open — or the day —, and a routine’s note after the routine.
- Where it goes. Into the vault’s Inbox folder (Settings → Notes & filing), under a name taken from your question — in a conversation a skill started, from the skill and the note that was open, or the day. A note that is already there is never touched: the new one gets the next free name. Plainva opens it right away.
- Who wrote it. The first line says it in words — an answer by Plainva AI, with the model, the time and your question. The note’s properties say the same for other tools:
generated, with the model and the time. Nothing marks the note as checked; that stays yours to do — see OKF. - What it rests on. Below the answer, Sources lists what the run really used. Plainva writes this list from its own record, not the model: the pages that were read and when, the searches and through which provider, and your notes that went along or were read. The property
sourcescarries the pages, the notes and the pictures, but not the searches. - Addresses. Every web address the model wrote into its answer is written so that nothing opens or loads it (
https[://]…), and an image from the web is never an image in the note. Only the pages under Sources are real links: addresses the run read with your leave. Links to your own notes stay links. - Rules. A kept answer inherits the privacy rules of what it rests on. If a note that went along with the conversation, or one the assistant read, is kept from the cloud or from the internet, the new note carries the same rule — written into the note itself where its folder would allow more. So an answer that a model on this device made from a private note does not reach a cloud as a note either.
In a shared workspace its members can read a note — and so can the readers of a publication that covers the folder. There Plainva asks every time, with the note’s name and the folder: Keep as a note writes it, Don’t keep leaves it.
What may leave this vault
Some notes should never reach a cloud provider. A rule can sit in a note’s frontmatter:
plainva:
ai:
cloud: deny
or, for a whole folder, in Settings → AI in this vault, in the card What may leave this vault (a table of the questions To a provider and Onto the internet, each answered allowed, not allowed or as the folder above; while the vault keeps the internet off, that column says not in effect – the internet is off on this device), which writes the rules to .agent/policy.yml. A note kept from the cloud contributes nothing — neither text nor title — and links to it in other notes are withheld. That holds however a link spells the note — by its file’s name, its title or a path —, and where two notes share a name and one of them is kept back, a link by that bare name is withheld as well; write the folder into the link to name the one you mean. Models on this device stay allowed. Encrypted workspaces keep the cloud off unless you allow it there. The exact format is in the File Format Reference. In a database, too, such a note counts for nothing: where the assistant reads one, a count, a sum or a list over links (a rollup, or the notes that link to an entry) leaves it out. What the assistant finds and changes follows the same view: its search does not find a note by the text of a withheld link, and a change that would touch one is refused. What is withheld are links: a name that merely stands in the text of a note that may go, without a link, is part of that note and goes with it. Two things remain: the order of search results and the time a search takes follow the index of the whole note, so a program or model can confirm single words of a withheld link’s text with targeted searches — never the content of the note kept back.
A conversation led with a model on this device stays on this device. Such a model may read what no cloud may, and a conversation always goes to its model as a whole — so once a conversation has run here, switching it to a cloud model sends nothing: a notice says so, your words stay in the input, and New conversation starts one with the model you chose. The other way round works: a conversation with a cloud model can go on with a model on this device.
A picture belongs to the notes that show it: one that a note kept from the cloud embeds goes to no cloud model either (see Explaining an image above).
A second rule, web: deny — not allowed under Onto the internet in the settings — keeps a note or a folder out of every conversation that may use the internet.
On an iPhone or iPad, the same two rules decide which note titles Siri and Shortcuts may find, once you have switched that on: a note kept from the cloud or from web access is never named to them. See “Siri and Shortcuts” in The mobile app.
History and usage
Conversations stay on this device, per vault — never in the vault and never synced. Keep conversations decides how long; you can delete single conversations in the list or all of a vault at once. A conversation you delete while an answer is still on its way stays deleted. When the vault is reloaded while an answer is on its way — after Rebuild index, for example —, the answer stops; what it wrote so far stays in its conversation in the list. Usage this month sums up the tokens per provider and model for this vault, from the newest 500 entries — one for every run of a conversation, one for every pass of the short summaries and of the search by meaning, one for every transcription and for every summary of a transcript, and one each time Plainva learns from a conversation, fills a database column with AI or turns a filter in words into filter rules. Each line says how much of the input the provider read from its cache, where it reports one, and the cost, where it publishes prices. Below the models, Saved by choosing and shortening estimates the tokens this month’s requests did not carry because Plainva sent a section of a note or its key sentences — verbatim or as a short summary — instead of the whole note: the rest of each such note, up to the most Plainva sends of one note. A version of a note counts at most once per conversation — with the system’s own model on this device, which gets every turn’s notes anew, once per turn —, and not at all where the model read the note itself while answering that question; a note that went whole saves nothing, nor does one only listed — by name or with a one-line summary —, nor a request that reached no model. The saving shows in dollars only where every one of those models has a price; a model on this device counts as costing nothing.
Limits of the beta
- On the desktop the AI runs in the main window only.
- On the phone an answer comes only while the app is open.
- Out of the box the assistant changes no note itself: a change to a note and the transcript of a voice note are suggestions you accept or decline, something new is a draft until you press Create, an e-mail or an appointment a draft until you send or save it yourself, and a rename, a move or a deletion waits for your yes; in a comment thread it writes a reply beside the note, never text in it. An answer becomes a note only when you press Keep as a note; Plainva writes it then, not the assistant. Where you switch on Apply directly or a folder allows Change without asking, it writes changes and new notes directly, and so does a script it starts in that conversation — with a version before and Undo (see Apply directly).
Feedback on the beta goes to the project’s discussions on GitHub: Feedback on the AI (Beta) in the settings starts one.