Run a personal AI butler on Telegram that asks before acting

A non-developer has run Francis, a Telegram butler built on 27 n8n automations with Mistral models, since May 2026; it handles mail, calendar and research but asks on Telegram before any outside action.

Done withn8ninTelegramPostgresqlMistral AI· Nov 17, 2026

What they did

Matteo Möller, posting on the n8n community forum as CivilisedGandhi, built a personal assistant called Francis that has run around the clock on a small Hetzner server in Docker since May 2026. He sends it jobs by text or voice on Telegram: find and answer emails, add calendar entries, research on the web, file recipes. A language model, mostly from Mistral, maps each request to one of 19 tools, and each tool is its own fixed n8n workflow rather than free-form chat. Unprompted, it sends a morning digest of emails that need him plus the server's health, and a weekly review on Sundays. Memory sits in Postgres with pgvector. He says he is not a developer: language models wrote most of the code while he supplied the ideas, research and testing.

How it went

It has been in daily use for about six months and now runs 27 automations. Every action that reaches the outside world, such as sending a mail, creating an event or filing a ticket, first arrives as a Telegram card with buttons, and the workflow only continues on his approval. A shared error workflow pings him on Telegram when any workflow fails, and a monitor checks the external connections. He gives no figures on cost or time saved.

Worth knowing

The design choice that matters is routing to fixed workflows with an approval button before anything leaves the system, instead of letting a chatbot act freely. Every paid model call goes through one sub-workflow that logs its cost, which keeps spending visible. The full build write-up is in German on the author's site.

Try it

With n8n· 5 parts to fill in

Build me a personal assistant in n8n that I talk to through [Telegram or my chat app]. Use [model provider] only to classify each message into one of these tools: [list of jobs, such as search mail, reply to mail, add calendar event, web research, save note]. Make each tool its own workflow with fixed steps. Before any step that sends, books or creates something outside the system, send me a message with Approve and Cancel buttons and wait for my answer. Store things I ask it to remember in [database]. Log the cost of every paid model call in one shared sub-workflow, and alert me in chat whenever a workflow fails. Every morning at [time], send me the emails that need a reply.

Square brackets are the parts to fill in. The prompt is ours; the story is theirs.

Where they said it

Related