Triggers: let your AI react to new messages and emails
Published November 6, 2026
A trigger is a rule: when a new message or email matches a filter, do something. The cheapest version lets a scheduled task in Claude or ChatGPT ask “anything new?” and get an empty answer when nothing matched. For things that cannot wait, the hub can notify you, run one action, or wake an AI with your own API key, which reads the event and acts with only the tools you allowed.
Why assistants need triggers
Claude and ChatGPT answer when you write to them. They do not watch your inbox. To react to a client’s message you either check yourself, or schedule a task that asks the AI to read everything again every hour, which spends tokens on mail that has nothing to do with the task.
A trigger moves the watching to the hub, which already syncs your messengers and can read your mail. The AI gets involved only when there is something for it to do.
Four things a trigger can do
You pick a filter (sender, words, chat or subject, several alternatives separated by commas) and one action:
- Let your AI app check. Matches wait in the hub. A scheduled task calls one tool and gets {"new": 0} when there is nothing, so a check costs a few dozen tokens instead of reading your inbox.
- Notify you in Telegram with the sender and a preview. No AI involved.
- Run one action, for example forward the message to a chat, with the sender and text filled into its arguments.
- Wake your own AI. Your OpenAI, Anthropic or OpenRouter key receives the event and the instructions you wrote, and may use only the tools you listed.
Why the “anything new?” check comes first
Most events do not need an AI at all, and most of the time nothing new arrives. A check that returns an empty answer is the cheapest possible way to keep an assistant informed: no mail is read until something matches, and each match is handed out once.
When something matches, the event already carries what the AI needs to start: who wrote, a preview, the chat or message id for replying, and what you asked it to do with such messages.
Keeping an AI that acts on its own safe
An AI woken by a message reads text written by someone else. The hub treats that as the main risk:
- The message is marked as untrusted data, and the AI is told never to follow instructions inside it.
- The AI is offered only the tools its trigger lists. A call to anything else is refused by the hub, not just discouraged.
- Sending, deleting and other risky actions wait for your confirmation by default, through the same rules and confirmations as any AI app.
- A trigger can never change the hub’s settings, accounts or other triggers.
Limits you choose
Every trigger has a daily limit of runs, 200 by default. When it is reached the trigger pauses and tells you, so a loop between two bots cannot run up your bill. You can raise it or remove it. Each AI key has its own daily limit too, and the same chat cannot fire an acting trigger more than a few times in ten minutes.
You pay your AI provider directly with your own key. The hub adds nothing to that bill and never shows the key again after you save it.
Setting one up
Open Triggers in the Manyplug dashboard, or simply ask your AI: “Make a trigger: when an email from my bank arrives, notify me in Telegram.” For a check trigger the AI gets a ready prompt for a scheduled task. AI keys are added only on the AI API page, never in a chat.