qhkm/zeptoclaw — reverse-engineered prompt

Build me a fast, small, local first personal AI assistant that runs as a single Rust command line app. I want to be able to install it on my laptop, a tiny edge box, or a cheap VPS, go through a simple onboarding flow for API keys and workspace setup, then chat with it from the terminal, stream replies live, run batch prompts from a file, and optionally expose it as a gateway for things like Telegram, Slack, Discord, or webhooks.

It should feel secure by default, not like a toy. Give it built in tools, memory, support for multiple model providers, and let it do agent style tasks like web fetching and shell commands inside a sandbox with approval for risky actions. Add guardrails for prompt injection, secret leaks, bad shell commands, and suspicious network access. If there is a control panel, make that optional. If it makes sense, include a migration path for people coming from OpenClaw. Look up current docs online if you need to.