Questions, answered

Agent Babysitter — abs — is an open-source remote control for Claude Code. It watches a session running on your own machine, messages your phone on Telegram when a task finishes, and takes your reply straight back into that same live session. Screenshots, voice notes and Claude usage limits all work from the chat. It is one bash script plus Anthropic's official Telegram plugin, on macOS and Linux.

NOIt's your own private bot, and only your paired account can reach it. One honest caveat: Telegram itself sees the messages (they're not end-to-end encrypted), so don't use it for work where that matters.

NOUnpaired accounts are dropped before Claude ever sees them. See Security for exactly how the allowlist works.

YESTranscription and speech both run on your machine — no cloud speech vendor is involved. The audio still travels over Telegram, like any message.

YESOne bot per project, via profiles — abs --profile work. Each runs its own session and reports to the same phone without fighting over messages.

YESTelegram polls outbound, so no public IP, open port, or webhook is needed. Run it in tmux on a VPS or home box and close the laptop.

NOIt runs anywhere Claude Code runs — a laptop, over SSH, or a headless box. A server is just one option, not a requirement.

PARTLYMessages are read between turns, so you can halt the plan and steer the next step from your phone, but not interrupt the step that is already running.

Linux and macOS, both tested — including macOS's ancient bash 3.2, with no Homebrew or GNU coreutils needed. Windows via WSL should work but is untested; reports welcome.

YESAll of it. abs prompt opens a tabbed page over every layer — the persona, the wording of each remote command, this project's CLAUDE.md, your global one, and what Claude has remembered. Edit what is yours, reset any part back to the shipped text, and the bridge mechanics stay read-only so personalising it can never break it.

NOFinished work waits in the working tree until you say so, because you may still want to change it. Nothing is committed to tidy up and nothing is pushed as a side effect of another task — push, deploy and tag each ask one plain question and wait for your yes.

YESFour ship — abs itself, plus ceo, cto and friend — and you can write your own. You pick one at launch and it shows in the status bar. A persona changes the voice and the priorities; it can never change the safety rules, which are appended after it.

Claude Code does the work; this is the part that lets you leave. On its own, Claude Code needs you at the terminal to see that a task finished and to say what happens next. Agent Babysitter reports to your phone, carries your answer back into the same session rather than starting a new one, and adds the things that only matter when you are away — a written summary, voice, usage checks, and rails so a message from a phone cannot trigger something irreversible.

YESThat is the core of it. When a task finishes or the session is blocked on you, a message arrives with a written summary of what happened — not just "done". Reports go quiet on their own while you are typing at the terminal and resume when you step away, and abs quiet on mutes them without cutting anything inbound.

DEPENDSOnly if the machine stays awake. A sleeping laptop suspends the session and picks it up on wake; messages you sent meanwhile are pooled and offered back, so nothing is lost, but no work happens while it sleeps. For work that genuinely continues while you are out, run it over SSH in tmux on a desktop or a small VPS, or stop your laptop sleeping on mains power.

YESAttach an image in the chat and Claude reads it directly — a failing screen, a stack trace you photographed, a design to match. It becomes part of the instruction, exactly as if you had described it, with no copy-paste through the terminal.

YESSend a voice note and it is transcribed, then repeated back to you before anything is acted on, so a mis-heard word gets caught early. Replies come back spoken as well as written, with six voices to pick from by ear. Both directions run on your own machine — no cloud speech vendor.

YESMIT licensed, and the whole thing is one readable bash script plus two small Python helpers — no framework and no build step. The repository is public, and the docs link every feature straight to the line of source that implements it.

YESAgent Babysitter costs nothing and there is no account to make with us. You need a Claude Code subscription — the one you already have — and a free Telegram bot, which takes about a minute to create. Nothing runs on our servers, because there are none.

Your bot is private to you: pairing is confirmed with a PIN at the terminal, and the allowlist can only be changed there, so a message asking to be let in is dropped. Destructive commands are blocked when the instruction came from Telegram rather than the keyboard, and five control phrases can mute, halt or cut the bridge entirely. The honest limits are written down too — Telegram is not end-to-end encrypted. The full threat model.