Skip to main content
Choose New task, pick a local project or cloud workspace above the composer, then the agent, model, and reasoning effort. Cloud tasks can also start from integrations, the CLI, or the API. The rail on the far left switches between your computers and your cloud workspaces; the panel next to it lists that place’s projects or tasks. Each computer shows its operating system and initials; in Twill Desktop the computer you are on comes first, and a dot on the others shows whether they are online (green) or offline. A badge on a computer or workspace counts the requests waiting for you there. Switching back to a workspace reopens the task you last had open in it. ⌥⌘↑ / ↓ (CtrlAlt on Windows and Linux) moves through the rail, and ⌘B folds the panel away while the rail stays. To reorder cloud workspaces, drag them in the rail. To hide one, right-click it and choose Hide from sidebar, or Reorder and hide… to see every workspace: drag rows (or focus a handle and use the arrow keys) and click the eye to hide or show one. Hidden workspaces also leave the new task picker; you stay a member, and Search still finds them. The arrangement is saved in this browser; Reset restores the default.

​
Runs and follow-ups

Each message starts a run. Follow-ups reuse the task’s harness and checkout; model and reasoning effort can change within that harness. Messages sent while the agent is busy are queued. Stop cancels a run without undoing its file changes. The agent’s questions and plan approvals wait for your answer in every permission mode. Integrations post questions back to their thread; open the Twill task when an interactive approval needs you. Open the task’s context panel to inspect branches, pull requests, subagents (with their prompt, activity, and answer), and background work.

​
Commands

Type / in the composer to list the commands the selected harness runs natively — the same ones you would type in its terminal. Add an argument after the command:
  • /plan uses the harness’s planning mode. Approve the plan in the task chat or reply with changes; follow-ups stay in planning mode until approval, or until you remove the planning pill.
  • /goal (Claude Code and Codex) asks the harness to keep working toward an objective across passes. It applies to the message you send it with.
  • Claude Code also runs its headless built-ins (/compact, /context, /usage, /recap, /init, /security-review, /mcp) plus every skill, custom command (.claude/commands), plugin command, and MCP prompt in the environment. Commands that need an interactive terminal are left out.
  • Codex supports /compact, /review (with optional instructions), /init, and each installed skill by name.
  • OpenCode supports /compact, its configured commands (init, review, .opencode/command/*.md, MCP prompts), and each installed skill by name.
Before a task’s first run, the menu shows the built-ins (and, for local tasks, what the runner discovered on your computer). Once a run starts, the harness reports the real inventory and the menu updates. An unrecognized command is sent as plain text; an unsupported one reports a capability error. Commands are separate from routing triggers, which select a cloud agent and can precede them: @twill /ultra /plan redesign the export pipeline.

​
Background work

Agents can leave long commands, monitors, or subagents running in the background and end their turn. When that work finishes, the agent picks up where it left off. Claude Code, cloud and local. The task is free for your next message right away while the background work keeps running. When it finishes, the agent starts a new turn on its own to report the result. A message you send in the meantime is answered by the same session, and the background work carries on. Other agents. The task stays running with the status Waiting on background task. Sending a message ends the wait: Twill stops what is still running, keeps the agent’s answer, and starts your message.
Background work is bounded to 30 minutes after the agent’s turn; Twill then stops what is still running. Stop cancels a turn together with its background work. On local tasks, background work needs an up-to-date desktop app and stops if the app quits or the task moves to the cloud.

​
Results

  • Local: changes land in your folder. Review them in Changes or your editor, and preview the app on a local port. See Work locally.
  • Cloud: implementation tasks normally create or update a pull request with a verification summary and any test output, screenshots, recordings, or preview links. Cloud agents are instructed to run checks and report what they couldn’t verify. Failed CI on a linked PR can trigger an automatic fix.
A Twill pull request with verification output and screenshots
Evidence reflects the environment where the check ran. After a handoff, ask for checks again if the destination’s services differ.