Docs

Agent

Chat

Use the chat composer to give the agent context, attach files and layers, pick a model, and follow its work.

Chat is where you give the agent work and follow what it does. The composer at the bottom holds your message, the context you attach, the mode, and the model.

Send a message

Press Cmd + L to focus the input. Type what you want, then press Enter to send. The placeholder changes with the mode, for example Describe what's going wrong in Debug.

Lead with the outcome, then the constraints:

Add a settings page that matches the Settings frame on the canvas.
Reuse the form components in components/ui. Run the typecheck when done.

While the agent works, the send button becomes Stop Generation. Messages you send during a run wait in a queue above the composer (2 Queued). For each queued message you can Edit, Delete, Force Send it now, or pick the model it runs with. Drag messages to reorder them.

To change something you already sent, click your message, edit it, and send it again.

Mention files, folders, and layers

Type @ to add context as a badge in your message:

MentionWhat it adds
FilesA file from your project
FoldersA folder from your project
LayersA layer or frame on the design canvas
PagesA whole canvas page
ProjectsA recent project to open, when no folder is open

Layer and page mentions point the agent at the exact canvas node, so it reads the real design instead of guessing from your words.

To grab something on screen, click Pick an element in the composer and choose an element in the built-in browser or a layer on the canvas. The button shows only while a browser tab or the canvas is open.

Use skills and commands

Type / to open skills and commands.

  • Skills are reusable instructions, each saved as a SKILL.md file in its own folder under .claude/skills/, .velork/skills/, or .codex/skills/, in your project or your home folder. Give each one a name (and ideally a description) in its frontmatter. Pick a skill to add it to your message; the agent also sees your skills and loads one on its own when it fits. When a project skill and a personal skill share a name, the project one wins.
  • /compact summarizes the conversation so far to free up room in the context window.

Attach files and photos

Click + and choose Attach to add files, photos, or folders. Photos go in as images; files and folders become mentions. You can also paste or drag images into the composer.

Click the microphone to Dictate instead of typing. Dictation needs you to be signed in and to allow microphone access.

The + menu

The + menu holds the composer's settings:

ItemWhat it does
AttachFiles, photos, and folders
ThinkingMore reasoning before the agent acts, on models that support it. For Claude models, Thinking lives in the model picker instead, next to the effort level.
AutonomousA supervisor keeps the agent going until the goal is met, sending follow-ups and answering its questions for you
Computer UseLets the agent control your Mac's screen (Develop and Debug only)
Cloud computerLets the agent use a cloud computer instead of your screen (Develop and Debug only)
Ask / Approve for me / Full accessHow shell commands are approved. This changes your Settings → Permissions, not only this chat. See Permissions.

Pick a mode and a model

The chip after + shows the current mode. Click it to switch, or press Shift + Tab to cycle. See Modes.

Open the model picker with Cmd + /. You can switch models in the middle of a chat. On models that offer it, the picker also sets the effort level. Connect more models under Settings → Models.

Shortcuts in the composer

ActionShortcut
Focus chat inputCmd + L
Cycle modesShift + Tab
Open model pickerCmd + /
Toggle ThinkingCmd + Shift + T
Toggle AutonomousCmd + Shift + A
Toggle Computer UseCmd + Shift + U

The mode, model, and toggle shortcuts work while the chat has focus.

Follow the agent's work

The agent's tool calls appear in the chat as cards: commands with their output, file edits as diffs, canvas changes, browser steps, and a todo list for longer jobs. Choose how much of this you see in Settings → General → Chat detail:

  • Detailed shows full cards for commands, edits, and tools.
  • Compact shows one-line rows.
  • Minimal shows only the agent's updates and answers, with one status line while it works.

The ring under the composer shows how much of the context window the chat uses. When the chat has cost money, the amount billed shows next to it. Hover the ring for the breakdown.

Undo the agent's changes

Each of your messages has a Restore Checkpoint button. Use it to roll files back to how they were before that message. If you change your mind, click Redo under the message to bring the changes back.

To see everything the agent changed, open Review Changes with Cmd + Option + 4.

Commit and open pull requests

Ask the agent to commit, push, or open a pull request in plain words. It commits only what belongs to the change and opens pull requests as drafts unless you say otherwise.

Shape how the agent answers

  • Settings → Instructions holds text added to every new chat, such as your stack or preferred tone. Use @ there to mention files and folders.
  • Settings → General → Caveman mode makes answers shorter: Caveman keeps every fact in fewer words, Ultra uses fragments.
  • Settings → General → Notifications tells you when an agent finishes or needs you. It is on by default.
  • AGENTS.md, CLAUDE.md, or CONTEXT.md at your project root is read in every chat for that project. See Overview.

Next steps

Modes

Pick the right mode for the job

Sessions

Run several chats at once

Permissions

Decide what runs without asking

Keyboard shortcuts

Full shortcut reference

Still stuck?

Ask us anything, or see what changed in the latest release.

Changelog