The front desk

A private chat where I text my own AI, and things on my laptop actually change.

livesince Aug 2026in the repo: Telegram assistant

LLife OSfront desk

task renew car registration by Nov 3012:40

Filed. Due Nov 30. I'll remind you on Nov 23.

fast path · no AI · 0.3 s

12:40

voice note · 0:0612:44

🎙 “what did the plumber say about the water heater?”On Oct 3 you noted: replace the anode rod before winter, about $150.

searched the archive · 1 hit · 9 s

12:44

Draft to the plumber is ready.

SendEdit

12:45

Sample with invented data. The real one looks the same.

What it does

When I send a message, Claude Code answers from inside my Life OS folder, with my tasks, tools and house already loaded. It can run a script, switch a light or read a file.
When the message is a short command ("task …", "note …", "c2 done"), it is handled instantly with no AI call at all. Free, fast, and immune to pasted text trying to trick it.
When I send a voice note, it transcribes it and shows me what it heard, so a mis-hearing is visible before anything happens.
When it drafts an email, it posts a Send button. Nothing goes out until I tap it.
When any other part of the system has something to say, it says it here, so there is one voice instead of ten apps.

1 · Hears
A small program on the laptop polls Telegram. It answers exactly one person and silently drops everyone else.
2 · Thinks
Short commands take a fast path with no model. Everything else starts or resumes a Claude Code session with today's deadlines and a search of past conversations in front of it.
3 · Does
The reply streams back to the chat. Long jobs detach and report back later, with a "still working" ping so silence never means "stuck".

Two days of silence
Every reply failed because a Windows command line can only be 8,191 characters long. The first "fixed" was declared too early: I tested a fresh session while the live one resumed an old one. Now the rule is to test the exact path the live system takes.
The Send button that never sent
It drafted emails for weeks through a connector that turned out to be read-only. It now sends through my own Gmail login, and a dead login fails loudly in the chat.
The allow-list that wasn't a wall
I proved that a tool missing from the bot's allowed list still runs in headless mode. Only explicit deny rules stop a tool, so that is what guards the one thing it must never do: move money.

The call that mattered

The AI drafts, I tap. Anything that leaves the house waits for a human.

Built with

Node.js, no dependencies · Telegram Bot API · Claude Code (headless) · ElevenLabs speech-to-text · Gmail API