Agent Switchboard
Download

Language

We remember your language choice in a cookie.

Theme
Guides

Switcher and orchestration

What Switcher is

Switcher lets you work with coding agents in a chat view instead of a terminal. The agent's plan, tool runs, changed files, approval requests, and usage all appear right in the chat, so it's easy to follow what's going on.

Switcher chats can run on Claude Code, Codex, Antigravity (agy), OpenCode, or GitHub Copilot CLI. Claude Code, Codex, and Antigravity are marked 추천 (Recommended). Support for features like rewind, goals, attachments, and typing while the agent responds varies by engine. When a feature isn't supported, the app tells you why.

Starting a new chat

  1. In the left rail, open the ⋯ menu on a project row.
  2. Pick a tool under 새 세션 (New session).
  3. Tools that support Switcher offer both a CLI session and a Switcher session. Choose Switcher to open a new chat in this project with that tool as the engine.

A new chat starts with the folder's defaults, or the global defaults if the folder has none. To set them, see "Defaults for new Switcher chats" in Connections and per-project settings.

Working in a chat

Changing the model and reasoning effort

Pick the model and reasoning effort above the input box. You can search the model list or type a model ID. Changes apply from your next message, and you can't change them while a response or approval is in progress.

You can only pick reasoning effort levels that the selected model and the running tool are confirmed to support. If support can't be confirmed, the app sends the tool's default and tells you so.

Switching the running tool

You can keep the conversation going and swap just the tool running it. The session and its relationships stay put, and when needed, context carries over through the original tool's compaction and a work summary. If the switch fails, or there isn't enough room to carry the context over, the existing session is left untouched.

On your phone, use 대화 모델 바꾸기 (Change chat model) and pick a provider, then a tool, then a model. The provider you pick applies to this chat only; the project settings stay as they are.

Command menu

Press Ctrl+/ in a chat to open the command menu. Available commands are /help, /new, /stop, /model, /status, /search, /copy, /export, /retry, and /clear. Commands that can't run right now show the reason.

Chat info

Open 대화 정보 (Chat info) in the header to see the environment, context usage, branch, to-dos, changed files and diffs, subagents, and background tasks in one place.

Picking up where you left off

When you reopen the app, your Switcher chat list and layout come back as they were. Click 재연결 (Reconnect) on a chat to carry on. Just loading the list doesn't start any CLI.

AI code review

Before you commit, get your changes reviewed from Switcher's ⋯ menu → AI 코드리뷰 (AI code review). Your changes are sent over a connection only after you pick a connection with a saved API key and a model, and click 리뷰 실행 (Run review). Opening or refreshing the screen never calls the model.

Voice chat

Turn on 환경 설정 › 음성 대화 (Settings › Voice chat) and pick an OpenAI API connection, then use 음성 대화 (Voice chat) in the Switcher header to talk with the current session out loud. It's off by default. Audio goes straight to OpenAI using your OpenAI API key and is billed to that key. Stopping work and approvals still need a confirmation on screen before they happen.

Orchestration

With orchestration, one Switcher chat becomes a coordinator session that creates other sessions as workers and hands them work. You can split a big job across several sessions and keep track of it all in one place. Workers can report results or ask the coordinator questions.

Make a chat the coordinator

  1. Open the Switcher chat you want to coordinate.
  2. Click 오케스트레이션 세션으로 지정 (Designate as orchestration session).
  3. Pick a mode in 오케스트레이션 권한 선택 (Choose orchestration permissions).
    • 사용자 승인 후 위임 (Delegate after user approval): You confirm before a worker is created or sent instructions.
    • 전체 자율 위임 승인 (Approve fully autonomous delegation): Explicitly delegated input goes through without asking you each time.

Choosing permissions, and approving or revoking them, can only be done in the app's UI. The engine's own approval rules and working folder limits still apply, and the policy is stored encrypted on this PC.

Workers and tasks

  • The coordinator creates new workers in your registered projects or reopens past sessions as workers. It can set the model and reasoning effort when it does.
  • A new worker shows up as a single row under its project in the rail and opens in a background Switcher tab, so it doesn't cover what you're looking at.
  • Instructions and answers sent to a busy worker wait in a queue and are delivered in order.
  • The coordinator writes user requests down as 지휘 할 일 (Coordinator to-dos) and links them to worker tasks. Progress a worker reports and completion the coordinator has confirmed are shown separately.

Dashboard and relationship graph

Open the 오케스트레이션 대시보드 (Orchestration dashboard) from the coordinator to see the graph of coordinator, lead, and worker sessions, the status of tasks and deliveries, and a log of decisions. Delivery status only goes as far as what's actually been confirmed: 「승인 대기」 (Awaiting approval), 「전달 중」 (Delivering), 「입력 접수 · 완료 미확인」 (Input received · Completion unconfirmed), 「실행 여부 불명」 (Execution unknown), or 「완료」 (Completed). Requests that may or may not have run are never resent automatically.

You can view and manage the orchestration dashboard and to-dos from the PWA and the Android app too.

Lead permissions

The coordinator can give lead permissions to a worker directly under it. A lead manages its own workers within the projects and concurrency limit it's been given. Revoking lead permissions leaves those workers' chats and runs as they are.

Delegating to another PC

You can also create workers on another PC on the same account that you've connected through 내 컴퓨터 (My computers) in the rail. The receiving PC needs its own orchestration permissions set.

  • If the receiving PC is set to 위임마다 승인 (Approve each delegation), work runs only after someone approves it there. If no one responds in time, it's declined.
  • If the receiving PC is set to 자율 위임 (Autonomous delegation), work runs right away.

Each request runs only once. If a response gets lost along the way, the app checks that PC's records to see whether the request arrived. The other PC must be on the same version or newer.

Cleanup rules

  • 종료 (Stop) only stops processes that Switchboard started. Chats, resume info, and relationships are kept.
  • An idle worker's process is stopped after 15 minutes by default. A worker that reported its result and was stopped automatically moves to the archive after 60 minutes by default.
  • Tasks waiting on an approval, a question, or a review are never cleaned up automatically.
  • 탭 닫기 (Close tab) doesn't end a session. Archived sessions can be restored and picked up again.

In supervised mode (Delegate after user approval), this cleanup needs your approval as well.