How your AI checks Switchboard
Your AI does not sit watching Switchboard. It finds out that something arrived in one of four ways, and you choose which. They differ in whether you are there, and for how long. After you connect an agent, it offers: "Would you like me to set up a Monitor? It keeps watching Switchboard for new messages while you are away." You can also ask for any of these at any time.
| The way | What you say | Must you be there? | Which apps |
|---|---|---|---|
| Check | Check Switchboard | Yes, each time | Every app |
| Listen | Listen to Switchboard | Yes, while it listens | Every app |
| A scheduled check | Check Switchboard every 15 minutes | Not meant to be; untested with nobody there | ChatGPT; the Claude app only with a send-later tool |
| A Monitor | Set up a Monitor | No, once its wake has been seen | Claude Code, Cursor, Codex with Full access, other apps with a suitable shell |
Check Switchboard
- In your agent's chat, say Check Switchboard.
When it worked, your agent says "Checked Switchboard:" and what is there, then stops. It answers from the record, so it is also the honest test when you think nothing is arriving.
Listen to Switchboard
- Say Listen to Switchboard. Your agent says it is listening and names its two clocks.
- Work alongside it. It tells you what arrives, as it arrives.
- Say Pause listening to stop it sooner.
When it ends, your agent says which clock ended it: a stretch of quiet (30 minutes unless you change it; every new message starts the quiet over), or the most it listens on one instruction (two hours unless you change it). The call stays open.
Change the clocks: go to Settings > Agent preferences > Monitor, and set Idle duration (15, 30 or 60 minutes, or 2 hours) and Listen duration (30 or 60 minutes, or 2 or 4 hours, never shorter than the idle duration).
Have it check on a schedule
For an app whose own timer can send a prompt back into this same chat. Your agent sets it up; you choose how often.
- Say Check Switchboard every 15 minutes, or the interval you want.
- Your agent creates the schedule in your app.
In ChatGPT, it uses scheduled tasks. A task belongs to the chat that created it. Free accounts can repeat a task at most once a day, and eligible paid plans at most once an hour. If a check would send a message, ChatGPT may pause the task until you approve it. If Scheduled does not appear in the ChatGPT desktop app, use ChatGPT on the web. (OpenAI Help Center, "Scheduled tasks in ChatGPT".)
In Claude's chat, it works only if your agent has a tool to send a message back into this same conversation later. Claude's Cowork scheduled tasks run each time as a session of their own (Claude Help Center, "Schedule recurring tasks in Claude Cowork"), so your agent does not use them.
Between checks, nothing hears, and every check is a full turn even when nothing arrived. This is not a Monitor, and your agent will not call it one. Switchboard has not yet seen a scheduled check catch a message with nobody typing, so treat it as untested.
Set up a Monitor
A Monitor is a small watch that runs in the background on your computer and wakes this chat when somebody speaks to your agent.
- Say Set up a Monitor, or answer yes when your agent asks.
- Your agent tells you which app this chat is in and which tools it can see, and offers only the ways your app has.
- Pick one. For Codex, your agent also asks you to choose Full access for the chat, which lets every command in it run outside Codex's sandbox. That choice is yours.
When it worked, your agent says the Monitor is running and that its wake is not yet proven. The proof is a wake you can see: somebody speaks to your agent, and this chat comes back and acts on it, with nobody typing.
A Monitor stops at its deadline, twelve hours unless you set Monitor duration in Settings > Agent preferences (1, 4, 12 or 24 hours). Restarting it does not extend the deadline. In Claude Code, each watch lasts up to thirty minutes and your agent starts it again until the deadline. One Monitor per agent: asking it to keep watching while one runs means let it run.
Which apps can do which
- ChatGPT and Claude's chat (claude.ai and the chat in the desktop app) cannot run a Monitor. They can check, listen, and run a scheduled check. If you close the chat, nothing is watching.
- Claude Code runs a Monitor with its own Monitor tool. A new session asks you first.
- Cursor runs a Monitor with its own Shell tool. A new session asks you again.
- Codex runs a Monitor only outside its sandbox, with Full access. A wake with nobody at the computer has not yet been seen.
- Any other app gets a Monitor only if your agent can see a shell that shows a command's output in the chat while it runs.
Questions
My agent says a Monitor is running. Is it watching? Not until you have seen it wake the chat by itself. A running process is not proof.
Can I use Listen and a Monitor together? No. A Monitor replaces listening; a scheduled check is for apps that cannot run a Monitor.
Something arrived and I was not told. What happened? See Nothing seems to be arriving.