OpenClaw is an open-source AI agent framework with admirable ambitions. It supports 20+ messaging platforms, 35+ LLM providers, a hub-and-spoke gateway (Node.js), an agent runtime (separate language, naturally), a three-tier skills override system, built-in RAG pipelines, multi-agent routing, cron scheduling, browser automation, and voice wake-word detection. It has 400,000+ lines of code, 1,800+ open issues, and requires at least 1 GB of RAM to breathe. The official docs recommend a $600 Mac mini as a minimum viable host. Cold-start on a modest CPU: ~500 seconds. Security researchers have called it "a security nightmare dressed up as a daydream."

To be fair, if you need to orchestrate 47 Telegram bots across 12 LLM backends with a RAG pipeline and overnight autonomous tasks, OpenClaw is probably fine.

Most people just want to talk to Claude on their phone.

NOC is for those people.

How it works

User sends a message to the bot
First message must be the auth passphrase, otherwise the bot replies "not authenticated"
Once authenticated, messages are piped to claude --resume <session_id> via stdin
claude stdout is streamed back as the reply (message is edited live as tokens arrive)
Files uploaded to the bot are forwarded to Claude; files Claude writes to the output dir are sent back
Sessions are scoped per chat and refresh daily at a configurable hour (default: 5am)

Setup

cp config.example.yaml config.yaml
# edit config.yaml with your values

Config

Key	Description
`tg.key`	Telegram bot token
`auth.passphrase`	Passphrase required to authenticate each session
`session.refresh_hour`	Hour (local time, 24h) when sessions reset

Deploy

make deploy

Builds the release binary and installs a systemd --user service. Remote deployment via make deploy-hera.

Logs

journalctl --user -u noc -f