NerdAlertAI is a self-hosted agent with an actual personality — and a permission system that means it. It reads your email, watches your network, briefs you on Telegram, and asks before it touches anything. Your hardware. Your models. Your rules.
$ GitHub — coming soon
$ GitHub — coming soon
NerdAlertAI runs on your machine and talks to the models you choose — Claude over API, Mistral on your own GPU through Ollama, or anything OpenRouter serves. The core loop is small, locked, and boring on purpose. Everything interesting is a module: turn it off and the agent behaves as if it never existed.
Six levels, L0–L5, set by you in config. Every tool declares its floor. The agent literally cannot reach above the rung you've granted.
Sending an email? Writing to GitHub? A card appears with the exact action. Nothing executes until you press Approve — enforced in the broker, not the prompt.
Sherman, Kenny, Brett, Toshi, Darius, Brooke, Bridget — the full Nerd Alert roster. Voice changes. Privilege never does.
Point it at your own Ollama box and your conversations never leave the LAN. Cloud models are an option, not a requirement.
Chat at your desk, get briefed on your phone. Critical alerts arrive immediately; routine ones wait until morning.
Credentials live in your OS keychain via a loopback-only setup page. No tokens in .env, no keys in chat — the agent refuses them on principle.
Every agent in the Nerd Alert cartoon roster ships ready to run. Same tools, same rules, very different bedside manner.
TypeScript and Node. One repo, no Docker required, no account to create anywhere.
A local-only page collects API keys and credentials straight into your OS keychain. Nothing is written to disk in plaintext.
Sherman is on duty by default. Swap operators any time from the web UI — no restart, no reconfiguration.