Per-project routing
Add a folder, then choose a connection (API key, gateway, or subscription login), model, and reasoning effort for each tool. Those settings apply only in that folder. Folders without their own settings use the global defaults.
- The 「도구 설정」 (Tool settings) cards on the project home show each tool's connection, model, reasoning effort, and scope at a glance. Click a card to change it.
- Codex gets a folder-specific profile, Claude Code gets
.claude/settings.local.json, and Pi, OpenCode, and Qwen Code each get their own project settings file. - A connection test checks the endpoint, authentication, and whether the model you picked actually works. If the model isn't available, nothing is applied.
- The original settings are backed up before the first change, and every change after that adds another backup. Once applied, the file is read back to verify it, and anything that looks wrong is rolled back.
Switcher chat
Work with coding agents in a chat inside the app instead of a terminal. Claude Code, Codex, and Antigravity CLI are the recommended engines; OpenCode and GitHub Copilot CLI work too.
- Reasoning, tool calls, file changes, approval requests, and usage all show up in a single chat timeline.
- Reopen the app and your chat list and layout come back as you left them. Reconnect and the chat picks up where it stopped.
- Choose the model and reasoning effort per chat. With 「이 대화에만 적용」 (Apply to this chat only), one chat can use a different connection while the project settings stay as they are.
- The command menu, history search, Markdown copy and export, and AI code review are all right there in the chat.
Pick up on your phone
See your PC's terminals and Switcher chats on your phone and type into them directly. Connect with the PWA or the Android app.
- Phone connection is off by default. Even when it's on, your PC only makes outbound connections to the relay server and never opens a port.
- The relay server passes ciphertext along without storing it. Every connection agrees on a fresh session key from ephemeral X25519 keys generated on both sides.
- The first time you connect, scan the QR code, then check that the 6-digit number in the PC's approval dialog matches the one on your phone.
- Notifications are encrypted with a separate key for each device, so neither the relay server nor the push service can read them.
Orchestration
For big jobs, a coordinator session splits the work across several worker sessions. You follow progress and results in one place.
- Each user request becomes a shared to-do linked to a worker task. A worker reporting success doesn't close the task; it stays open until the coordinator confirms the result.
- The dashboard shows each worker's reported progress and blockers, tasks waiting on an answer, and tasks that have stalled.
- A graph of coordinator, lead, and worker sessions shows how a large run is going at a glance. You can also open the original session on a remote PC.
- Pick the model and reasoning effort when you create or resume a worker. Idle workers are stopped and archived safely.
Multiple PCs and usage
Open another PC signed in to the same Google account from 「내 컴퓨터」 (My computers) in the rail, and work with its projects and sessions from the PC you're on.
- By default, adding another PC for the first time means confirming a 6-digit number in an approval dialog.
- The app tallies token usage by folder, tool, model, and date from your local CLI session logs. This is off by default.
- Turn on totals across PCs to sync token counts, never chat content, encrypted with your account's statistics key. The server never has that key.
- Enter prices for each model and connection to see estimated API costs as well.
Supported tools
Some tools get full connection management; others are only installed, detected, and launched. Available on Windows x64, macOS (Apple Silicon), and Linux x64.
OAuth tokens stay with each CLI. Agent Switchboard never reads or copies them.