What's shipped, what's next, and what's on the wall of monitors marked "soon." Dates are intentionally absent — features land when they're right, and the core loop never moves.
Trust ladder with elevation requests, permission broker with argument validation, approval cards on every model path, an audit trail tying every action to its approval, seven personalities, Gmail / Calendar / GitHub modules, SOC monitoring, scheduled jobs, memory engine, Telegram bridge, web search and weather, OS-keychain credential storage, and local-model support through Ollama.
A unified transport layer so any OpenAI-compatible endpoint plugs in with its own key — entered through /setup, stored in the keychain, and held to the same per-model trust ceiling local models already run under today. Weaker models simply get a shorter leash.
Ongoing work to make sure smaller local models never claim an action happened unless it actually did. Safety isn't just about blocking bad actions — it's about never lying about good ones.
Image, audio, and video generation as a fully optional module. Disabled by default, like everything else.
Speech in, speech out, on a Raspberry Pi. Talk to Sherman from the kitchen. Sherman has, of course, already heard you coming.
A smarter memory layer that curates itself — promoting what matters, letting trivia decay, and keeping the agent's picture of your world current without manual housekeeping.
Expanded dashboard coverage and after-action reports — pull an incident's alerts from across the stack and hand you a structured timeline instead of nine open tabs.
A pre-imaged Raspberry Pi that boots straight into NerdAlertAI for people who'd rather unbox an operator than compile one.