What it does

Your chats

Pin, file and tidy your chats, see which need you, write and queue messages, answer its questions with a tap, and keep a long chat going.

Your chats are in the sidebar. Point at one and press ⋯, or right-click it, for what you can do with it. On a phone, press and hold.

Your list

From the top, the sidebar shows:

Pinned apps. App pages you pinned, as a row of app tiles. Each tile has its app's icon. Point at one to see where it's from, like "Page of the Tally app". Right-click a tile, or press and hold it, for Unpin and the app's details.
Needs you. Chats waiting for you to allow or answer something stay here until you do.
Pinned. Chats you keep at the top.
Folders. Your own groups of chats.
Everything else, by when. Today, Yesterday, Previous 7 days, Previous 30 days, then one group per month.
Archived, at the end.

You only see a part once there's something in it.

What each chat is doing

A small mark beside a chat says what's happening in it:

The pearl: it's working.
An amber dot: it needs you.
A dot and a bolder title: something new happened while you were elsewhere, like a reply that finished or a message from a chat app. Open the chat and the mark goes, on your phone too.
A red dot: the last reply didn't finish.

The menu beside Chats shows All chats, only what's New, or only chats From chat apps.

Pin a chat

Choose Pin to keep a chat at the top. There's no limit. Drag pinned chats to put them in your own order, or use Move up and Move down. Archiving a chat unpins it.

Folders

Choose Move to → New folder…. Give the folder a name, and pick a mark and a colour. Then:

Drag chats onto a folder, or choose Move to and the folder.
Press a folder's name to fold or unfold it.
Press ⋯ beside a folder to Edit or Remove it. Removing a folder puts its chats back in your list. It deletes nothing.

Folders only sort your chats. They don't change how your assistant answers.

Many at once

Hold Ctrl or ⇧ and click chats to select them, or choose Select from a chat's menu. A bar at the bottom offers Pin, Move to, Archive and Delete. Done or Esc ends it.

On a phone

Swipe a chat to the right to pin it, or to the left to archive it. Undo puts it back.

Tidy up

Once a few chats haven't been touched for a month, a note at the end of the list offers Archive them. It leaves out anything pinned, in a folder, working, waiting for you or new. Nothing is archived until you press it, and Undo brings them back. Not now hides the note.

⌥↑ and ⌥↓ go to the chat above or below. Double-click a title to rename it.

Writing a message

↵ sends and ⇧↵ starts a new line. You can start typing anywhere in the chat: the words go in the message box.

↑ in the empty box brings back what you sent, newest first: this chat's messages, then what you sent lately in other chats. ↓ walks forward again. Change it and send, or keep going.
What you were writing stays in each chat when you go to another, and after a restart. Signing out on this device clears it.
Point at a message of yours and press Copy to take its words.

While it works

Esc or the stop button stops the reply. It ends there at once with a Stopped mark (and how long it ran), and you can write your next message straight away: if the provider is still winding down, it waits a moment and then goes.

While it works, the words above the reply say what it's doing after each step: reading a command's output, looking at a page, checking an edit. Beside them, a quiet tally: how long this step has taken, and what the reply has written so far, rolling up as it grows (12s · 42.3k tokens). The clock starts again with each step, so a long job reads step by step rather than as one big timer; each finished step keeps its own time on its row. Once a long reply is done, Worked 12m · 412k tokens among its buttons says what the whole of it took. The ring beside the mode picker keeps saying how full the chat is, and pressing it while a message runs shows what it has written and read so far.

Long jobs

Your assistant can take as many steps as a job needs: booking something across a few websites, tidying a big folder. A message you're watching runs until it's done, with no limit on steps, time or reading. It only stops by itself when it's plainly getting nowhere: the very same step coming back with the very same answer, again and again in quick succession. Checking on something that takes a while, failing tests on the way to fixing them, or a search that finds nothing are all work, never a reason to stop. It's told first, twice, and tries another way; only if it carries on does it pause, with Carry on to pick up where it was.

The reply then ends with one sentence saying why it paused. Press Carry on and it picks up exactly where it stopped, with everything it had done. Or say what to do differently.

Want it to check in sooner? In Settings → Usage → Long turns, turn on Pause long turns to check in and set your own limits: steps (starts at 100), minutes (30) and fresh tokens in millions (2), meaning what a message reads and writes that the provider hadn't already cached. A chat then pauses at the first one it reaches, with Carry on. Over your monthly budget, it checks in sooner still, but it never stops you.

Routines and background tasks have their own room (200 steps, an hour), since nobody is there to press Carry on. That doesn't change with the switch.

You don't have to wait to say what's next. Write it and press ↵: it waits above the message box, Sends when Conch is done, and goes by itself the moment the reply is over. Anything more you send meanwhile joins it.

The pencil takes it back into the box to change it, and the cross doesn't send it.
If you stop the reply, or it fails, the message comes back to the box instead of going.

When it asks you something

Sometimes your assistant needs your choice before it can go on: which day suits you, how you'd like to talk, how many people are coming. It asks with a card in the chat, and the reply waits for you.

Tap an answer. A question with one set of options goes as soon as you choose. Anything more has Send.
Pick a day from the row of days, or Pick a date for another. On a phone the row slides sideways for more days.
Pick a time from the few it suggests, or Another time… for any other. Numbers and a few words work the same way.
Something else… lets you write your own answer.
Or type your answer in the message box. While a question waits, it says Answer above, or type it here.
Skip lets your assistant carry on with its best guess. It tells you what it assumed.

On a keyboard, 1 to 6 choose an option and ↵ sends.

Once you answer, the card folds to one line with your answer. Stopping the reply skips the question, and so does Conch restarting while it waits.

A waiting question shows in the sidebar like anything else that needs you. With notifications on, your phone hears that your assistant has a question. In Talk, your assistant reads the question out, and what you say back answers it.

Routines, background tasks and chats from your chat apps never stop to ask, because nobody is there to answer. Your assistant picks the sensible choice and says which.

What to say next

Under the latest reply, up to three small buttons hold what you might well say next: “Make it shorter”, “Add Ada to the invite”. Press one and it's sent, exactly as it reads, as if you'd typed it. The chat's model and settings stay as they are, and anything you were writing in the message box stays there.

They come from two places:

Your assistant offers them when the next step is clear. Most replies have none.
Conch reads the reply itself. Under a table of numbers, Show it as a chart draws it beside the chat.

They go as soon as anything newer is in the chat: a message from you, here or on another device, or a new reply. You'll never see them:

while something else is waiting for you in that reply, like a question or a card to connect an app, so there's one thing at a time;
from your assistant after the chat has read something from outside, like a web page or an email, because those words could be someone else's (Conch's own still show);
in routines, background tasks or chats from your chat apps, where nobody is there to press them.

Rename

Choose Rename, type the new title and press ↵. Esc keeps the old one. A title you write is never replaced by the one Conch suggests.

Find what you archived

Archived, after the last chat in the sidebar, shows how many chats are in the archive. It opens Archived chats, most recently archived first. You can also type "archived" in CtrlK.

From there:

Choose a chat to read it. A note at the top says it's archived, with Unarchive.
Unarchive puts a chat back in your list, where it was.
The bin deletes it, after asking.
With many chats, a box finds one by name. ⋯ has Unarchive all and Delete all….

Delete

Choose Delete, from the list or from Archived chats, and confirm. The chat and what was attached only to it are removed from Conch. Anything your assistant remembered from it stays in memory, where you can forget it too.

Its plan, as it works

When something takes a few steps, like tidying a folder or fixing a bug, your assistant writes its plan into the reply as a short checklist. You see where it's up to while it works: finished steps get a tick, the one it's doing now shows what it's doing (“Running the tests”), and the rest wait their turn. A long plan shows the steps around the work, with Show all for the rest.

When the reply ends, the plan folds to one line, like Plan · 5 of 5 done. Press it to see the steps again. A reply you stopped keeps its plan as far as it got.

Claude Code and Codex keep a plan of their own, and Conch draws it. Other providers that can use Conch's tools get the same checklist. A model that can only chat doesn't keep one.

Plan only

In Plan only mode, your assistant looks around and plans, but changes nothing. With Claude Code, when the plan is ready it's shown to you in full, with two buttons:

Start begins the work.
Keep planning stays in plan mode, and puts you back in the message box to say what to change.

Long chats

A chat can go on as long as you like, with any model. Each model reads only so much at once, so when a chat grows past that, Conch writes a short summary of its start. The model reads the summary instead of the oldest messages, and carries on.

A quiet line in the chat shows where the summary starts: Earlier messages are summarised for the model's name. Press it to read exactly what the model keeps.
Every message stays in the chat for you. Only what the model reads gets shorter.
Before the start of a chat is summarised, Conch learns what you said there, as the memory tidy-up does.
If the model still says the chat is too long, Conch summarises more and sends your message again by itself. Only if that isn't enough does the chat offer a model that reads more at once, or a new chat.

How full is it? The ring beside the mode picker fills as the chat grows, with the share beside it (35%). It measures against the model you chose: a model that reads a million tokens fills ten times slower than one that reads 100k. It turns amber from three quarters and red from nine tenths. Press it to see how much the model reads each time, out of how much, and Compact now.

Type /compact, or press Compact now, to summarise the start now. Add what matters most to you, and the summary keeps it: /compact the garden plan. In CtrlK, Summarise the start of this chat does the same. Claude Code and Codex keep their own memory of the chat, so there /compact goes to them, and they summarise it their own way.

Good to know

Chats from your chat apps archive the same way. A new message from there brings the chat back.
Routines, tasks and pinned apps keep their runs on their own pages, so they're never in this list or the archive.
Archiving, pins and folders are kept with your chats, so they're in your backups too.
In CtrlK, type a folder's name to open it, or Pin this chat, Move this chat to… or New folder.