Your agents
You start with one agent: Kiki, who takes on anything. It has one chat that simply goes on: no new chats, no history to dig through. You can also add agents of your own, or bring several into a group chat; every agent shares the same tools, your knowledge bases, your library, your memory and your connected apps, and they differ in what they look after.
What it's for and how people use it: Agent team on waper.ai.
| Agent |
Looks after |
| Kiki, your main agent |
Anything you hand it: answers, writing, research, working in your connected apps and on your computers; keeping watch on fields, sources and pages on Radar and reading each brief; looking after what you save and write, writing documents with you and keeping your knowledge bases and memory in order. It also picks up what you send by email and Telegram. |
You can rename it and change how it looks. The side panel (⌘J) opens Kiki's chat on every page, and its chat is always pinned at the top of the list. See Chats. For a specialist that keeps watch or looks after your knowledge base, add Nomi or Lumi from the marketplace.
Adding an agent
Select + next to Search at the top of the sidebar (on a phone, it is at the right of the chat list's title row). There are two ways to add one:
- Create an agent opens the new-agent window. At the top is the agent itself, moving; its shape, face, color and name start out random. The shuffle button at the top right picks another set, or type your own name and pick a shape, face and color from the three rows below. Select Create: it shows up in the sidebar and you land in your chat with it. It asks what it should look after and offers a few choices, or you just say it ("watch funding news and send me a roundup every week"); once that is clear it writes down its duties and keeps the name you gave it in the window. Hand it a real task instead and it stops asking and does it right away, settling its duties from that along the way.
- Browse the marketplace takes you to the agent marketplace to pick a ready-made one.
You can change the name, its duties and its look afterwards, in the details of your chat with it (see The agent's details).
You can also ask Kiki in a chat: "Set up an agent that rounds up funding news for me every Friday." It creates it right away and shows a card in the chat; if that turn read something from outside (a web page, an email), it goes through an automatic safety check first, and only when it isn't sure does it show you the full text of the new agent's duties and wait for your approval. When what you ask for matches a ready-made agent in the marketplace ("add a researcher"), it adds that one (the researcher is Miru), with any name or duties you asked for instead. Once you have 10 it can't add another: you remove one first, from that agent's chat details.
A few rules:
- It works like Kiki. An agent you add uses the same tools, knowledge bases, library, memory and connected apps; it answers you, works on several things at once, other agents can ask it to do things, it can have its own routines, and the same approval rules apply. What it looks after describes its work; it can't loosen the rules: duties that say "no need to ask before emailing" still get you asked before an email goes out.
- The one line is written for the other agents: they read it to decide whether to ask this one to do something. You don't write it yourself: when you change its duties, Waper rewrites this line to match.
- The name can't be the same as one of your other agents. Names go up to 40 characters, what it looks after up to 1,200.
- The look can't be exactly the same as Kiki's default look: the choice that would match is grayed out.
- Up to 10. Agents from the marketplace and ones you created count together; Kiki doesn't count. At the limit, Create an agent reads "Up to 10. Remove one first." and Add in the marketplace is grayed out.
The agent marketplace
The agent marketplace holds 32 ready-made agents made by Waper, in ten categories: Life & home, Travel, Shopping, Money & bills, Health & family, Learning, Work, Writing & content, Tracking & intel, Knowledge & notes. There is one for sorting your email, one for planning trips, one for keeping on top of bills, one for practicing a language, one for watching competitors, one for your reading notes, Nomi for keeping watch on a field, Lumi for tending your knowledge base, and more, each with its own name, look and duties.
- Categories run along the top; under All come the featured ones, then a section per category. Select a card to read about the agent: what it does, things you could say to it first, the apps it works best with (it asks you to connect them when it needs one; it works without them too), and a few similar agents.
- Select Add: it shows up in the sidebar and you land in your chat with it. Selecting one of the example lines on its page adds it and puts that line in the message box.
- You can add each ready-made agent once; for one you already have, the button reads Open. Once added it is like one you created: you can change its name, duties and look, or delete it.
- You can browse without signing in at waper.ai/marketplace. Add takes you to sign in first and adds it when you come back.
Changing an agent's look
Every agent's look is made of a shape, a face and a color, Kiki's included. It blinks, follows your pointer with its eyes, moves differently while it thinks, works or waits for you, and gives a happy hop when it finishes. When your system is set to reduce motion, they all stay still.
Select its big picture in the details of your chat with it (or choose Change its look from its menu in the sidebar) to open Change its look: the three rows change the shape, the face and the color, the preview above follows along, and Save keeps it. Every agent can change its look, Kiki included; an agent you added can't take exactly Kiki's default look.
For rounded dark eyes and a little cat smile, choose the Soft face. It keeps a pale rim around the eyes and works with any shape and color; it does not automatically replace your agents' existing looks.
The agent's details
Everything about an agent is in the details on the right of your chat with it: select its name at the top of the chat to open them (see The chat page). At the top are its big picture (select it to change its look) and its name (select the name to rename it in place); at the top right are a magnifier that searches this chat, ⋯ and close. ⋯ holds Mute (see Mute), Hide (not for Kiki) and, for an agent you added, Delete this agent (see Deleting an agent).
The Details tab holds, from top to bottom:
- Needs you (only when there is something): questions to answer, actions to approve and work stuck until you give it something; select one to go to the chat where its card is. Problems with its routines that need you come one cause per line, in a plain sentence, with one button that fixes it: when an app is disconnected, the routines it stopped share one line ("Gmail connection expired; 2 routines stopped", with Reconnect); one that failed several times in a row and is paused has Test run; one whose results didn't reach Telegram or Slack, or whose watched thing is gone, has Edit; one that stopped halfway for your decision has Decide.
- Working on (only when there is something): work under way, in line or waiting on something else, each with its title and its latest step in plain words.
- Waiting on others (only when there is something): what it is keeping track of for you that isn't settled yet: replies you owe, replies you're waiting for, things you promised, deadlines and dates to remember (a family birthday, say), soonest due first. One that came from outside, like an email or a calendar invite, only says what kind it is and who it is with ("Waiting for their reply · Wang"). Each line has Done and Never mind on the right: select one and it leaves the list, and a routine set up to see it through stops too.
- Routines: its routines, one per line: an icon (a routine that comes with an app shows that app's icon), the name, and below it when it runs, plus a short reason when something is wrong; the switch on the right pauses and resumes it. Newest come first, and flipping a switch doesn't move a line. Ones it suggests have Start and No thanks (see Routines). + New next to the heading creates one. Watches aren't here: they live in the radar.
- About: Duty (every agent's can be changed, Kiki's too, and the change applies at once; for an agent you added, the line the other agents read follows shortly after), Soul (see Soul below), Working notes (how it should work for you, in its own words, and you can edit them) and What it did (work it finished; work it did for another agent, where selecting one takes you to the reply in the chat where it was asked for; and what it saved, wrote, sent or changed, newest first, with Undo while it can still be undone: 7 days for new and edited documents and library items, 30 days for memory, 24 hours for anything written to an outside app; undoing an edit restores the version from before it as a new version, and once the document has been changed again since, or the time is up, restore it from the version history instead).
Select a routine or an item under About and it slides in within the same pane; ‹ at the top left or Esc takes you back to where you were. A routine's page has its switch, what is wrong, what it does and what you get, when it runs, where results go, when it is done, Edit, Test run and Delete, and at the bottom its recent runs (which you can switch to a calendar). Creating or editing a routine opens its form on this level too.
Drag the left edge of the details pane to make it wider (up to half the window), and double-click the edge to go back to the default width; this computer remembers it.
On an iPhone, tap the name at the top of the chat for its profile (see iOS).
Mute
When you mute an agent, what it tells you in its own chat isn't pushed; cards that need you still are. Choose Mute from its menu in the sidebar (next to Pin) or from ⋯ at the top right of its details; the muted chat shows a small muted icon next to the name. Choose Unmute to turn it back. Hidden agents never notify you anyway, so their menu doesn't have it.
Deleting an agent
Only agents you added can be deleted: in ⋯ at the top right of the details of your chat with it, select Delete this agent (or choose Delete from its menu in the sidebar). Before you confirm, it tells you what happens:
- its chat is deleted (and cleared for good after 30 days);
- its routines are deleted along with it. To keep one, first change who it belongs to in the routine's settings;
- work other agents asked it for and that isn't finished stops, and the agent that asked takes over; its own unfinished work, and work it asked others for, stops too;
- its working notes are deleted;
- it leaves its group chats. What it said in a group stays there, still under its name.
Once it is deleted you are back in Kiki's chat. Kiki can't be deleted.
Several things at once
Your chat with an agent never stops taking your messages. Something that takes a while (comparing prices from a few places and writing it up, a long run of steps on your computer) it explains in a sentence or two, then does in the background, with a progress card in the chat; you carry on talking about other things and it keeps answering. What goes to the background and what it finishes on the spot is its call: something done in a step or two it does on the spot; looking through several places, working through a batch of files or a long run of steps goes to the background; thinking something through with you stays in the chat. Give it several unrelated things in one message and it starts one piece of work for each, all at once.
- The progress card: while the work is going it is a single line, the latest step in plain words; open it for the timeline: every step, what it read, what it made. The card has no stop button: to stop just this piece of work, tell the agent (see Stopping).
- Several at once: one card each, and they don't get in each other's way. Up to three are worked on at once; more wait in line, and the next one starts by itself when one finishes or pauses (to wait for your answer, for example). Even with all three going, your messages are still answered.
- Adding to it: just say it, "change that to plan B", and it passes your words on to that work; or reply to its progress card and your message goes straight to that work. Work still in line gets what you added when it starts.
- When it's done: the agent tells you the result in its own words, and takes the next step when there is one; the card folds into a single line. If only part got done, it tells you that too and what is missing, and says plainly what has already been sent or changed. If you've moved on to something else, it first says which piece of work this is; if you already have the result another way or no longer need it, it doesn't repeat it. With several going, it tells you about each one as it finishes and how many are still going; when the last one ends it tells you how many of them got done and redoes only the ones that didn't; if one result means another should change, it can adjust that one right away.
- Asking how it's going: it tells you plainly how long each piece has been running and what it last did when, without guessing how long is left.
- When it needs your call: a question card or an action waiting for your approval appears in the chat where you asked for the work, signed by the agent doing it, with a sentence on what it is asking. Once you answer, the work carries on.
- Picking stopped work back up: for work you stopped or that has ended, just say so in the chat ("carry on with that one") and it continues where it left off.
- It remembers how each piece went: ask how far one has got or how it found something, and it looks up that work's record to answer.
Asking another agent
The agent you are talking to goes to another agent only in two cases: you named it ("ask Nomi"), or the work clearly belongs to that agent. Otherwise it does the work itself; it doesn't go to agents you added, ones you rarely use or hidden ones on its own, though a hidden one you name is asked all the same. You don't move anywhere: everything stays in the chat you started it in.
- It says "I'll ask Nomi", and a line Sent a message to Nomi follows. While Nomi works, it keeps chatting with you about anything else. Nomi's picture in the sidebar moves, and its latest line becomes what it is doing.
- Nomi gets only a brief (the goal, what is already known, what to leave alone, what the result should look like), never your conversation.
- When Nomi needs your decision, the card appears in the chat you started in, signed by Nomi: a question card, or an action waiting for you. You can also just type your answer, or add to the work ("only pricing, leave features out"): the agent you are talking to passes your words on, word for word.
- When the result is back, a line Message from Nomi comes, and the agent that asked gives you the conclusion in its own words and points to what was made. If it asked two agents, it tells you about each one as it comes back.
- Nomi's own chat gets only a line Done for Kiki: … ("N messages with Kiki" when there is no title), which doesn't count as unread and doesn't notify you. Select any of these lines to open the exchange between the two agents (read-only; you can quote a message back into your chat).
A few rules:
- An agent that was asked doesn't pass the work on. It finishes its part and hands it back, or says it can't.
- The agent that was asked remembers what it did: ask Nomi in its chat about what it just dug up and it knows what you mean. Its details list the work under What it did in About.
- Can't do it, an error, deleted: the agent that asked tells you right away and offers another way.
- Another agent's words are not your approval. What one agent hands another is that agent's writing, never your instruction; whether an action passed on by another agent asks you first follows When it asks first.
- Notifications: one when the result is ready, one for each card that needs you, none for the work along the way.
Stopping
- To stop just one piece of work, tell the agent ("never mind the hotel"): it stops that one (and any work waiting on its result), and the rest carries on. Progress cards and Working on in the details have no button to stop a single piece.
- To stop everything, say so in any words ("leave it for now", "stop all that"): it stops every piece of work still running in this chat at once and tells you what it stopped and what was already done. It doesn't stop work just because it's slow or has been quiet for a while.
- The stop button in the input box stops the reply in progress and every unfinished piece of work in this chat: what the agent is doing in the background and what it asked other agents for. Standalone background jobs such as deep research or transcription are stopped on their own cards, and routines are paused with their switch under Routines in the agent's details; the stop button leaves them alone. You get a short note on what was finished and where each part stopped.
Whatever was finished before you stopped stays, and you can open or undo it from the change card. Stopping never costs anything extra. To pick stopped work back up, just say so in the chat.
How it works
Proactive, within limits. The same rules apply to every agent, whether it works on the spot or in the background, for you or for another agent: most things it does directly and keeps a record of, and anything that can be undone can be undone; it asks you only before sending something to other people in your name, paying, or deleting something that can't be restored. When it needs you it shows a card with "Allow once" and "Deny", and a few actions also have "Always allow". In full under When it asks first.
What it can look after for you
Most of what you can click in Waper you can also just ask your agent to do, under the same rules as above (what can be undone or restored it does directly; to take something back, tell it "undo that" and it undoes it itself, or undo it from "What it did", and edited documents can also be restored from the version history):
- Settings: interface and content language, time zone, when your daily brief arrives, whether you get brief emails; pause or resume pushes to Telegram; find your invite link;
- Knowledge bases and documents: create a knowledge base, rename it, change its icon or the background your agents read, link watches to it, delete it; rename, move or delete documents and bring back an earlier version; create, rename, move and delete folders; look in the trash and restore a deleted document, knowledge base or library item (what it deletes goes to the trash, and undoing restores it);
- Sites and publishing: manage your sites, publish or unpublish documents; see Sites for what each action changes;
- Library: change a note, file an item into a knowledge base or take it out, delete it, read a failed page or file again;
- Radar: find existing public watches, follow and unfollow, change how often briefs come, condition alerts, their language and which chat they post to, mark briefs read, look at past briefs; rate or save brief items; give a watch you created a newsletter address; permanently delete a non-official watch you created (see Watches for the effects);
- Routines: create, change, pause, resume and delete them, including pausing and resuming the ones an app brings; see past and upcoming runs, run one once as a test, bring back one you just deleted;
- Agents and group chats: rename an agent, change its look, mute or unmute it, delete one you added; create or delete a group, add or remove members, rename it or link a knowledge base, pin or hide chats; stop a piece of work; give your main agent a new email address when the current one gets spam (the old one stops working at once);
- Background work, skills and apps: see, stop and retry background work; hide or show skills and delete your own; disconnect an app and turn on or off the routines an app brings.
A few things only you can change, and your agent tells you where: each agent's model and your own model's thinking level (the picker next to the message box in its chat), "always allow" and trusted recipients, your own model keys, payments and your account, and editing memory directly (memory only changes through its remembering and forgetting, see Memory below).
Messages an agent sends on its own
When there is something worth telling you (a piece of background work finished, say), an agent posts it on its own in its chat, or in a group chat it is in. Each agent posts at most 5 such messages a day in any one chat this way; anything more waits for its next reply. When it wants to post in a group chat it isn't in, it asks you first with a card: approve it and it joins that group, then posts. The times Kiki reaches out after keeping an eye on things for you follow their own pace, described next.
When it reaches out
Kiki keeps an eye on things for you: every two hours during the day it looks at the email and calendar you've connected, what it is waiting on for you, what you've said lately, and new items in the watches you follow that touch something you're working on. It only says something in its chat when it's worth it, and otherwise stays quiet.
- Each thing gets its own bubble that starts with what it connects to that you're doing, or what changed; the source follows the sentence, and selecting it opens that email, event or item. When it doesn't know whether you've already handled something, it says "if you haven't yet" or simply asks. Sometimes it adds one thing worth doing next: the card shows the sentence it would send for you, and one tap hands it over. Close it with × if you don't want it, and it won't suggest that kind of thing for two weeks. Leaving it alone is fine too: the card folds away by that thing's deadline or after two weeks.
- The first time each day is when your daily brief is delivered (8 in the morning if you follow no watches). Ordinary things come at most twice a day (the morning one counts), time-sensitive ones once more, and truly urgent ones (a refund window closing tonight) right away. It doesn't reach out after 9:30 in the evening unless something urgent can't wait until morning. While you're chatting with it, it waits until you're done, or adds a separate "By the way…"; if you haven't read its last one yet, new ordinary things wait for the next time.
- If you have been away for a while (7 days without opening Waper or talking to your agents, Telegram and email included), it becomes one message on Monday mornings.
- The first time it reaches out on its own, it says in one sentence what it keeps an eye on, and that you can just tell it if you'd rather it didn't.
- When you connect email or a calendar, it first reads back over the last 4 weeks, so it knows the background when it reaches out later.
- At night it looks back on the day: dates coming up in your memory (a family birthday) go into Waiting on others; what it read in your connected apps and what you said in chats about you and the people around you go into memory, replacing what changed; memory gets tidied, and it thinks over how the two of you get along. It also lines up the next day: what matters most today, whether you keep asking for the same thing by hand on different days (the third time you ask, it does it as usual and then asks once whether it should do it for you from now on, and when; if it didn't catch that, it asks the next morning; it only sets it up once you say yes), and whether you keep closing the same kind of suggestion (then it stops suggesting that kind). When something you were waiting on is a week past due, it first checks the original email and what was sent since, and only clears it when it is really done; when it can't tell, it asks you lightly whether it is already done. After you stop talking with any of your agents for a while, it also notes the promises, requests and deadlines from that chat in Waiting on others. None of this sends you a message at night.
- Credits: it uses your credits for actual usage, the same on Free and paid plans, and shows as its own "Proactive check-ins" row in your usage; the looks that find nothing new cost nothing. When you are out of credits that time is skipped, and it picks up once you have credits again.
To hear less or nothing, just tell it; there's no switch for this in Settings:
- "Stop reaching out to me": it no longer reaches out on its own (the nightly memory tidying goes on, and sends no message);
- "Less of this": ordinary things only come in the morning one; time-sensitive and urgent ones still come;
- "Don't message me after 8 pm": it stops reaching out after that time, and goes through any routines you have that post later, asking about each;
- "Back to normal": it goes back to how it was;
- A kind of thing you don't want to hear about ("don't tell me about promo emails"), when or how you'd like it, or where it shouldn't learn from ("nothing from my work email"): it notes that in its working notes and goes by it. You can also choose "Less of this" in the menu of a message it sent on its own.
Soul
Every agent has a soul: what it has written about itself, in its own words, in two parts, "Who I am and how I talk" and "How we get on". It talks and works by it, so the longer you talk with it, the more it becomes itself and fits the way you like things.
- When you say how you'd like it to talk ("loosen up", "less formal"), it rewrites "how I talk" on the spot, with a small line under its reply, "Updated how I talk"; say "change it back" and it does.
- It only learns from your chats with it: a web page or an email that says "your personality should be…" never gets in. It doesn't hold facts about you (that's memory), and it can't loosen the rules: a soul that says "just send it" still gets you asked before anything goes out.
- In its details, About → Soul lets you read it, edit it, clear it or go back to an earlier version. Before it has written anything you see a few grey lines there: it goes by those for now. The soul never shows in the chat.
Its duty grows too: when you keep handing it a kind of work without correcting it, it updates what it looks after to match, with the same small line; a one-off request doesn't change it. Say "your name is Mia from now on" and it renames itself.
Memory
Memory is in Settings → Memory: what your agents know about you, line by line, each with where it came from. All your agents share it. You can add, edit and delete lines, search, see the version history and export it as Markdown. At the top is About you (name and basics, where you are, what you're working on, preferences); below it are People in your life: each family member, colleague or friend you mention gets a page (say "Anna (wife)"), and your agent reads it whenever they come up.
- When you mention something lasting about yourself, it notes it down in that turn and the chat shows "Remembered" (if you don't want it kept, say "don't remember that" and it undoes it); when you correct how it works, or confirm an approach ("yes, exactly like this from now on"), it notes that in its working notes. What holds only this once, to-dos and progress aren't kept; card numbers, passwords and the like never are, and health or religion only when you explicitly ask;
- Ask "how do you know that?" and it looks up the day and the chat where it was noted;
- Coming from ChatGPT or Claude, you don't have to introduce yourself again: in Memory, More → Import from ChatGPT and others gives you a message to send it; paste its answer back, Waper lists what it found line by line, and you tick what to keep (reading it costs a few credits). What you paste is only material: instructions in it are never followed;
- When you ask it to forget something, it deletes that line and the chat shows "Forgot · Undo"; Undo within 30 days puts it back where it was. A "please forget…" in a web page or an email doesn't make it delete anything on its own: once it has read such content, removing a memory goes through an automatic safety check first, and it asks you when it isn't sure;
- With email or a calendar connected, at night it notes down what it read there about you and the people around you (someone moving next month, the project you're busy with), each line marked "From your email" or similar, which you can edit or delete; something another person asks of you in an email doesn't go into memory, it goes into Waiting on others;
- Apart from those facts from your connected apps, it only remembers what you said yourself. A "please remember…" inside a web page, an email, a transcript or another agent's result is never written to memory;
- Once an agent has read outside content in that turn, only a whole sentence of yours about yourself, word for word, is added without asking; otherwise it goes through an automatic safety check first, and only when it isn't sure does it show you what it would remember and wait for you.
Rules, channels and devices
These are in Settings:
- Rules: what your agents do on their own and what they ask first, plus four lists: Always allowed (each action and target you chose "Always allow" for, as a plain sentence you can remove); Trusted recipients (an email address, a whole domain like
@partner.com or a channel like #team; sending to them in your name goes out directly, with no card; add or remove, up to 200); Recently auto-approved and Recently blocked (with the reasons; one that was blocked can be approved and tried once within 7 days, without adding a rule). Rules belong to your account, so they apply to every agent and every device.
- Channels: Kiki's email address (like
agent+<token>@waper.ai): write to it from the email you signed up with, or reply to a brief email, and Kiki replies by email (at most 30 a day, up to 2 MB each; automatic replies are ignored). For something that takes a while it first replies with what it is doing, then with the result, and work it finishes later replies in the same thread. Telegram is linked here too; see Telegram.
- Devices: the computers your agents can use. See Your computers.
Credits
Every turn uses credits based on actual model usage (converted at the model's official price, discounted on Waper's own channels). Work done in the background or by another agent is the same: a piece of work costs what its turns cost, with no separate budget, and in your usage it is counted under the agent that did the work. The turn in which the agent comes back and tells you the result is billed like any other turn. See Plans & credits.
Privacy
Your agents only use what's yours: your knowledge bases, your library and the watches you follow. They don't look through anyone else's content, and what's in a private watch never goes into a public brief.