Muse Code, Grok Build and Pi join, and Claude workflows run in the chat
Muse Code, Grok Build and Pi are now built-in agents. Claude workflows run live in the chat, and you can stop them. Agents can hand you a command to run, Cmd+F searches a chat, and replies start sooner because Codex and Claude start while you type.
Muse Code, Grok Build and Pi are built in
All three install, update and sign in from setup or Settings › Providers. Muse Code never bills an API key unless you allow it, and you choose whether Meta may train on your chats. Grok Build runs on a SuperGrok or X Premium plan, or an xAI API key. Revert also rewinds Grok's and Pi's own memory when it can, and says so when it can't. If you added Pi as a community agent, its row offers the switch.
Claude workflows run live in the chat
When Claude starts a workflow, or you type ultracode, a card in the chat shows its steps and agents as they work, and Stop workflow ends the whole run. The Agents tab shows each agent's transcript and the result. When Claude shares a workflow's script, you can read it there and save it as a command. Settings › Threads sets how many agents Claude aims for.
Run the commands agents suggest
Shell commands in an agent's finished reply get a Run in terminal button, and you can send the output back to the agent. On Windows that's PowerShell commands only. Threadlines pastes risky ones, like deletes, force pushes or sudo, into the terminal for you to check instead of running them. Agents can also ask you to run a command they can't, such as a sign-in, and they get the output when it finishes.
Find in a chat with Cmd+F
Ctrl+F, or Cmd+F on a Mac, searches the whole conversation, including the parts of a long chat that aren't on screen, and shows a count like 3 of 12. Enter and Shift+Enter move between matches, and a folded message opens to show one. Find skips work steps. Find in chat is also in the command palette, and you can change its shortcut in Settings. It's off on phones.
Replies start sooner
Typing in a thread whose Codex or Claude agent has stopped now starts it in the background, so your message goes to an agent that's already running. You can turn that off in Settings › Threads. Messages also reach the agent faster, because the snapshot taken before each turn reuses the last one. On a big install, startup work that took about 11 seconds now takes 2, and chats open faster.
Also improved
- The browser panel opens links meant for a new tab and sign-in pop-ups. You can trust a local dev server's own HTTPS certificate, and agent screenshots fail less often.
- With two computers on a direct connection, not the relay, an agent's browser opens on the one you're using. Quitting only warns about work on this computer.
- Claude goals use the goal bar, steering reaches Claude during a long command, and its dev servers can run past 30 minutes. The Usage page says what used up your Claude limits.
- Status checks run fewer git commands. On Windows the app starts far fewer git and PowerShell processes, and a project's own git.exe never runs in place of git.
- Quitting never gets stuck. A Stop background work button ends what an agent left running, and Stop reaches Cursor, Antigravity and community agents instead of waiting 15 seconds.
- The access menu shows the level each agent really runs at, and Auto on Antigravity now accepts edits without asking. The app no longer crashes at startup for T3 Code users.