Chat and Agent mode

Start from home

Chat opens on a centered composer with seven creation modes underneath — Image, Video, Websites, Pages, Docs, Sheets, and Slides. Picking one opens that mode's page, with starter prompts that prefill the composer and wait for your edit; nothing sends until you do. The active mode shows as a dismissible chip in the composer, and moving between home and a mode keeps your draft and any staged attachments.

Conversations name themselves

A new conversation is named for what it's about, in the language you wrote, in a few words. The name writes itself into the sidebar once it's ready. Rename it any time, and your name stands.

Talk instead of typing

Describing a process out loud is quicker than typing it. Press the microphone beside Send, or ⌘ ⇧ Space (Ctrl+Shift+Space on Windows and Linux), and say what you want, as you'd explain it to a colleague. Your words appear in the message as you talk.

• Hold the microphone, or the shortcut, to talk for as long as you hold it. Letting go finishes.
• Tap it to talk hands-free, then:
  • Done (✓, or the shortcut again) keeps what you said in the message, ready to edit before you send.
  • Send (or Enter) sends it as soon as your last words are in.
  • ✕ (or Esc) throws the recording away and puts your message back the way it was.

A recording can run for up to 5 minutes. From 4:30 a countdown shows, and at 5:00 it stops and keeps what you said, so you never lose it.

The first time, Cheqq asks which language you'll speak. Choose it: what you say comes out far more accurately than when Cheqq has to guess. Change it any time under Settings → General → Voice input → Dictation language. Your browser then asks to use the microphone. If you said no earlier, allow it from the icon in your address bar, then try again.

If part of a recording can't be turned into text, say because you lost your connection, it stays on the page until you press Retry or Discard.

Your recording is turned into text, then deleted. It is never stored. Each recording uses about 2 AI credits.

Choose a mode

• Ask is read-only. Use it to explain, compare, summarize, or inspect without making changes.
• Plan researches what's available and returns an inline markdown plan for you to review before anything runs.
• Agent (the default) uses the tools and connected services available in the current context to complete work.

The composer reads left to right: the + menu, the model picker, then the Chat mode menu. Change mode from that menu before sending a message. With Agent selected, the same menu also holds Approvals — see Control tool approval.

Recover or branch a turn

Every accepted user message creates a checkpoint before Cheqq begins the turn. Use Restore or Fork from the divider immediately above that message:

• Restore returns the conversation and captured local state to the point before the message while retaining the abandoned future as a recoverable branch.
• Fork starts an independent conversation from that earlier prefix and, on supported surfaces, duplicates the Wapp, workflow, or App the conversation is about.
• Undo restore returns to the preserved future after the latest applicable restore.

External actions that already happened are not reversible. A partial checkpoint identifies local state Cheqq could not capture. See History and recovery for the full behavior and what each surface can restore.

Write an effective request

State the outcome first, then add constraints, the relevant Wapp or service, and how you'll judge success. Attach a file when the task depends on its contents.

For example:

Review the attached vendor list, identify missing contact details, and plan the updates. Do not change any records yet.

Each attached file can be as large as your plan allows: 5 MB on Free, 25 MB on Pro, and 100 MB on Business and Team. A file over the limit is refused in the composer, before it uploads.

Choose the model and its effort

The model picker applies to the next message, not the whole conversation, so you can start a thread on a fast model and move to a stronger one when the work gets harder without losing the history. Auto, at the top of the list, lets Cheqq choose.

Models that reason show an Effort row at the bottom of the picker: one dot per effort level the model supports, with its own default marked in violet. Cheqq uses that default unless you move it. Turn it down for a quick answer or up for a harder problem; picking the default again clears your choice. Each model keeps its own setting, and the row follows whichever model will answer — including the one Auto picks. It's hidden for models that don't reason.

Generate images and videos

Open the + menu beside the prompt and choose Create image or Create video to force that output. The model picker only lists admin-approved models that can produce the selected output, including each model's supported input-modality icons and current credit burn.

The picker shows the models you've curated. Manage models… (pinned at the bottom of the picker) opens the full approved catalogue grouped by provider, with search, modality, and pricing hints — tick the ones you want offered and the picker shows only those. Your selection follows your account; if you never curate, the picker shows the admin-featured set.

You don't have to use the menu — in ordinary Agent mode, Cheqq can decide to generate an image or video on its own whenever it judges that's the best response to your request, as long as the workspace has an approved model for it. Use the + menu when you want to guarantee that output or pick a specific model; otherwise just describe what you want.

In video mode, describe the scene, subject, camera movement, and desired outcome. You can attach an image for models that support image-to-video generation. Generated videos appear in the conversation with playback controls.

Every picture and clip Cheqq makes is kept in Artifacts, and its caption in the chat says so — Saved to Artifacts, with a link to it.

Video generation can take several minutes. The generation is priced by its actual duration and selected resolution/audio variant rather than as text tokens.

Edit an existing image or video

Generation can build on what's already in the conversation. Asking for a change to an image — "make it warmer", "remove the background" — edits the picture you're looking at rather than starting over.

In Image mode, the composer's Editing chip names the picture your next request will change, with its version. Clear it with its × to start a new picture instead. An edit is kept as that picture's next version, and its caption reads Version 2, Version 3, and so on. If an edit should have been a picture of its own, select Keep as its own artifact in its caption.

Videos can be edited too, on models that accept video input. Video mode gains an operation control — Generate, Extend, or Remix: extend continues a clip past its last frame, remix restyles it. Attach the clip you mean, or leave it off and the conversation's most recent video is edited. You can also just ask the agent — "extend this video by another scene" — from an ordinary Agent-mode turn. Edits are billed at the model's own video-input rate, and every edited clip records its lineage back to the original.

Work that keeps going without you

Agent mode runs on Cheqq, not in your browser tab. Once you send a request, the work carries on whether or not you're watching: close the tab, lose your connection, or put the laptop to sleep, and when you open the thread again the task is further along, with its thinking, tool calls, and delegated agents' transcripts all still there. An approval it was waiting on is still waiting, and so is a question it asked you. You can pick it up from another device.

When a task needs you

While a task works, the reply shows what it's doing and the composer says Add guidance while it works… — nothing else sits in your way. A strip appears above the composer only when the task needs something from you:

A task that runs out of AI credits shows a card instead of a strip — see Allowances and Resume.

When Cheqq asks you something, the questions come one at a time, with the answer it recommends marked Recommended. Pick an option by its letter, write your own under Something else, or Skip a question to leave it to Cheqq. Next moves on, and Send answers them all.

Allowances and Resume

A task works in execution windows of up to two hours and 1,000 steps.

• Time. A window that runs out of time having made progress opens the next one by itself, a few times over, so a long build isn't interrupted just because the clock ran.
• Steps. A window that uses all 1,000 steps always stops and waits for you. Steps are what a task spends, so only you can grant more.
• No progress. A task whose runs keep ending with nothing new to show — twice in a row is enough — stops and comes back to you, so a task going in circles is caught while you can still point it somewhere better.
• Lost connection. If a run loses its connection, the task waits a moment and tries again by itself. It only stops and says why after several losses in a row with no progress between them.
• Credits. A task that runs out of AI credits pauses and shows the way to get more — a credit pack on a paid plan, an upgrade on Free — with Resume beside it. Topping up doesn't restart the task; you decide when it continues. A message refused for credits shows the same card and is never retried on its own.

When a task is Paused, the strip says which limit it reached. Select Resume to open a fresh window.

A message you send to a paused or blocked task counts as resuming it: the task continues with your message as its next instruction.

Add guidance while it works

Type while a task is running and your message is held above the composer rather than sent at once. Until it goes, you can read it back, edit it, or remove it. It's sent as the current reply ends, so it shapes the work still to come, and it appears in the conversation as your own message.

When it can't wait, select Send now. The task abandons the step it's on and starts again from your message. It's still the same task: its window, its progress, and what it has already done all carry on, and being redirected doesn't count against it. Send now is also how a held message with files attached is sent.

If a delegated agent is partway through its part of the reply, Send now waits for it rather than throwing its work away: the held message reads Sends when the agent finishes, and goes the moment it does. Select Send anyway to redirect at once.

A held message keeps what it was about. If you wrote it while asking about something specific — see Ask about something specific — it's sent about that, even if you've picked something else since.

Stop a task

To end the task instead, use Stop. The task reads as stopped straight away: the strip says Stopping while the step already running finishes, and the work ends within moments. Stop can't be pressed twice, and an approval the task was waiting on is closed with it. Even when a tool doesn't answer, the task is let go within half a minute, so your next message doesn't wait behind it.

Control tool approval

Cheqq runs in Auto unless you choose otherwise: it goes ahead with tool calls without stopping to ask. What it changes inside Cheqq — in chat, Wapps, workflows, and artifacts — can be restored and undone.

To change it, open the Chat mode menu in the composer. With Agent selected, the menu's Approvals section has three options:

• Auto — run tool calls without asking, saved as your default.
• Manual — ask before tool calls run, saved as your default.
• Auto for this session — stop asking for now without changing your saved choice; it turns itself off when the browser session ends.

Turning on Auto shows a confirmation first. It applies to every tool call, including ones made by delegated subagents and the connected services they use. Changes inside Cheqq can be undone, but a message or post sent outside it cannot, and instructions planted in a page or email the agent reads could trigger actions you didn't ask for.

Emails the agent writes from your Gmail follow their own setting and wait for your review by default, whichever approval mode you're in — see Send email from your Gmail.

In Manual, sending a new message while an approval is still pending automatically denies it.

Read an approval card

When Manual is active, a proposed tool call appears as a card that leads with what it will do in a sentence — Delete 42 records — followed by what it costs if you agree, and buttons that name the action rather than reading "Approve" and "Deny". Shell commands render as commands. The raw arguments are still there under View details when that's the question you have.

Each sensitive step asks on its own. Approving a delegated agent's first action does not cover the rest of its run — the card says so, and the agent checks with you again before the next one. If you'd rather stop being asked for the rest of the session, Always allow on the card approves that step and turns on Auto for this session in one gesture. It appears only where that session override exists.

Approvals survive a reload

Every ask a run raises is kept. Reload the thread mid-pause and the open question is still there to answer, in place beside the call it was about. Asks that have already been settled stay in the conversation as receipts, each naming the step it was about, so a run that paused three times has a record of all three rather than only the last.

Send email from your Gmail

With Gmail connected in Integrations and allowed to send, the agent can email from your own address: new messages, or replies that land inside an existing Gmail thread. One request can produce several emails.

By default, nothing goes out until you've looked at it. Each email appears in the conversation as a card — To, Cc, Bcc, Subject, and the message — that you can edit in place, then Send or Discard. ⌘/Ctrl+Enter sends from inside the card. A sent email collapses to a one-line receipt with Open in Gmail, and the agent hears on your next message which emails you sent or discarded.

To let the agent send as soon as you ask, turn off Review emails before sending in Settings → Advanced. The same setting applies in Slack and Telegram.

If Gmail doesn't confirm a send, the card says the email may or may not have gone out and points you to your Sent folder, rather than guessing. Deleting a conversation deletes its email drafts.

Watch delegated work

When Cheqq delegates part of a request to a subagent — building out a workflow, or handing work to another Wapp's agent — it appears as one line rather than an open transcript: who worked, how many steps, how long it took, and how it ended. Open it to read the full trail, including the agent's thinking as it goes, for as long as you want it; it collapses itself a beat after the run finishes so watching it work doesn't cost you the rest of the turn.

Select Open alongside on the row to move the transcript into the side panel beside the conversation, where it gets the room to be read while the main agent keeps working. It shares that panel with document and app previews, so opening one swaps the other.

A delegated agent's transcript is kept with the task, so reloading the thread doesn't reduce it to a summary line. A delegation that fails reads as failed, with its error; one you stop reads as stopped.

Follow what a thread builds

When a conversation builds something, you watch it happen beside the chat:

• A workflow is drawn on a read-only canvas that updates each time the agent revises it. Every workflow the thread built can be previewed, not only the latest.
• A Wapp shows its real workspace, live.
• An App shows its preview, as it builds.
• A page, document, spreadsheet, or deck appears in the reply as a card marked Open beside chat; select it to read it in the pane. These live in Artifacts.

Everything the thread has built sits above the composer. With one thing, that's its chip; with more, it's the current one's chip and a count that opens the full list, Built in this conversation. Each entry shows its name and whether it's still building or its build failed. Select one to show it in the pane, or use its open action — Open in Builder, Open Wapp, Open in editor, or Open artifact — to open it on its own page in a new tab.

While the pane is open, the sidebar folds down to its rail to give it room, and comes back when you close the pane.

Opening something from the chat never moves the conversation. The thread stays where it is, and what it built opens where it lives.

If a thread's Wapp, App, or workflow is later deleted, the conversation still opens and still answers. Its chip reads Deleted.

Saved workflows

A workflow the agent saves shows in the reply as one line — Saved workflow and its name — with its steps a click away and a link into the editor. If a reply saves several, each gets its own line, and they stay after a reload. A Workflow ready for approval card, with Approve, appears only when the agent finishes with a workflow it didn't save.

Chat inside a Wapp

Every Wapp has its own chat, and it starts with up to three suggestions of what to ask. They lead with what needs you — a failed automation, then setup you haven't finished — and otherwise fit the Wapp: an empty one offers a first prompt to finish and What can you do here?; one with data offers a question about the table on show, a weekly summary, or the piece it's still missing, such as a dashboard or an automation. If you can only read the Wapp, you're offered questions only.

The chat knows which page you're on. The composer invites you to Ask about the collection you're looking at, and a record's panel has Ask about this record, which opens the chat with that record already named in the composer, ready for your question. On a phone, the bar along the bottom of the Wapp does the same for the page on show.

Inside a Wapp, chat can delegate to your other Wapps' agents too. The Wapp you're in keeps its tools directly on the agent; every other installed Wapp is available as a delegate, bounded by your role on that install.

A business system's Getting started steps happen here too. A step like Add your clients opens this chat with the step, and Cheqq asks one question at a time, brings in a spreadsheet you attach, or reads a photo or PDF you send, until the step is done.

You can also ask the agent to delete a Wapp you made. It tells you what will go to Trash — and anything that would stop the delete — before asking you to confirm, exactly as the Delete button does. The Wapp can be restored from Trash.

Ask about something specific

When you ask about something you've selected — an element in an artifact, a comment you handed to Cheqq, a part of an App you picked — the message keeps it. Your bubble reads Asking about and names it, and still does after a reload, so "make this violet" never loses what "this" was.

Follow tool activity

While the agent works, its thinking, narration, and tool calls stream into one timeline. Tool calls are described in plain language from what they're actually doing — "Searching the web for supplier reviews" rather than a raw tool name — and on models that expose their reasoning, the agent's thinking streams between them. The timeline highlights the current step while the turn runs and collapses to a one-line summary when it finishes; open it to read the full trail, which survives reloading the thread. Reloading while a task is running picks the timeline up from the step being worked on, rather than replaying the run from the beginning.

Every part of a turn speaks the same way. The main agent's work, a delegated agent's transcript, and calls reported after the fact all render as the same timeline, so what you're reading never depends on where in the turn a call happened to land. A step summarises what it found without being opened, and a result you do open is complete and copyable rather than truncated.

Approvals, email drafts, browser handoffs, plans, rich results, generated artifacts, and delegated-agent rows remain separate so an important interaction cannot disappear inside the timeline.

Search the web and look at a page

Cheqq can always search the web, read a public page in full, list the pages of a site, or crawl one when a task needs site-wide coverage. There's nothing to connect first.

It can also take a screenshot of a public web page when it needs to see what a page looks like rather than read what it says — a layout to follow, or the design of a site you pasted. It uses a browser on Cheqq's side, not yours, and can only reach public addresses. The screenshot is saved to the conversation like any other image, and the agent sees it from its next step. If a capture fails, the agent can try it again.

To have Cheqq work in your browser — signed in, on your tabs — use the Companion extension.

Reference a file with @

Type @ in any chat input to open the file picker: this thread's files first, grouped by kind with thumbnails and clip durations, then everything in your workspace Files. A pick becomes a chip on your message.

Referenced files are the explicit subject of that turn:

• In image mode, the referenced image is the one edited — rather than the most recent picture in the conversation.
• In video mode, the referenced clip is the one extended or remixed.
• In Agent mode, "this" and "these" mean exactly the files you attached.

A Wapp panel's # mention opens the same picker, so there is one way to reference a file everywhere. If a referenced file has since left the workspace, the message is refused by name instead of quietly falling back to something else.

Work from a plan

When Plan mode presents a plan, review its scope, summary, and any clarifying questions it asks first. Selecting Implement this plan switches the conversation to Agent mode and starts the work as a tracked task list; Dismiss leaves the plan in place without acting on it.

Share or delete a conversation

To show someone how a conversation went, share it as a replay — see Share and remix.

Deleting a conversation stops any work it still has running, then erases what it left behind: its messages, the requests and results of its task, every delegated agent's transcript, its approvals, and its email drafts. The credits it used stay on your usage record, without their content.

If a run stalls or returns an error, keep the conversation open and follow Troubleshooting.