Voice assistant that actually gets things done. Press a hotkey, speak, and Nudge handles the rest.
Press a hotkey, speak, release. Nudge transcribes your speech, classifies what you meant, and routes it to the right action: tasks, alarms, notes, clipboard, links. No wake words, no fixed commands. "Remind me to call mom tomorrow" becomes an alarm in under a second. Bring your own LLM, runs local-first.
Built at the OpenAI x Outskill AI Builders Hackathon (May 2026), evolved from an earlier local-first voice experiment into something that ships. It's a macOS-native assistant driven by a global hotkey — pip install nudge-ai, nudge setup, and it's listening.
~/.nudge/soul.md file teaches Nudge how you work — what "later" means, how you triage priority.fast (~200ms, free Groq tier), default (~500ms, balanced), offline (~2s, fully local via Ollama). Swap instantly.I kept losing thoughts mid-coding-session. The friction of opening a todo app was enough to lose the idea. Nudge removes that: hotkey, talk, done.
Part of a bigger picture: Hive for orchestration, Unplug for defense, Nudge for voice. Each one explores a different facet of making AI work locally.
Python, built on Hive (agent runtime, intent routing, STT/LLM providers), Typer + Rich (CLI), Groq Whisper (STT), Pydantic config. A transport-agnostic engine drives the same pipeline across hotkey, CLI, and HTTP, and reports per-stage timing — STT → intent → agent, end to end around 740ms on the default preset. Sessions log to ~/.nudge/logs as JSONL.
Hosted backend with auth and STT proxy, continuous listening, calendar integration, and deeper Hive agent integration so a single nudge can kick off an autonomous follow-up instead of a one-shot action.