Live

Sterling

Personal AI assistant and second brain — an Obsidian vault with an agent layer for calendar, briefings and coding work, plus a local voice front end.

ObsidianHermes AgentClaude CodeLiveKit

Motivation

I am moving from keeping everything in my head to writing everything down, and I use AI for personal admin and productivity. The second brain takes small decisions and reminders off my mind, so my head stays free for the decisions that matter. Sterling is the assistant on top of it. It acts as my secretary and as a critic that pushes back when I slack or over-engineer.

Project Overview

Sterling has two parts. The first is my Obsidian vault of about 1,400 notes with an agent layer around it: rules for how agents behave, a playbook that routes coding work through stages, and a squad of agent roles (researcher, implementer, tester, verifier, validator, pragmatist, DevOps) that the main assistant delegates to. The second is Sterling Voice, a voice front end I can talk to from my laptop or phone. Every day Sterling sends me a morning note, asks in the evening what I did, and reads out what is on tomorrow.

Key Features

  • Obsidian vault as the single knowledge base, backed up every night
  • Agent squad with separate roles for research, building, testing and review
  • Calendar on Radicale CalDAV, synced to the phone
  • Morning note, evening check-in and a spoken evening briefing
  • Sterling Voice: a local voice assistant in German and English
  • Runs on Hermes Agent and Claude Code from the same rules

Architecture

The vault is plain Markdown in Obsidian, versioned in git and pushed to a private remote every night. Hermes Agent runs the scheduled jobs and the chat, Claude Code is the second runtime for coding work, and both read the same rules from the vault. Its skills are small Python scripts for a CalDAV calendar on Radicale, daily briefings, transcription with faster-whisper, journal capture and source checking, and the tools repo behind them has about 400 automated tests. Sterling Voice runs on my laptop with a LiveKit server, speech-to-text through faster-whisper or Deepgram, Hermes as the brain, and local Piper voices or a cloud voice in German and English. It is served as an installable web app, so I can use it from my phone too.

Challenges Solved

The rules live in the vault and each runtime only translates them, so a new runtime does not mean new rules. Agents must not invent facts, so reminders and briefings come from scripts that read the calendar, not from a model's guess. Whoever builds a change does not review it. Verification is its own role, on a different model where possible.

About this Repository

This repository is private, so there is no public source to browse. The description above is what this page publishes about the project.