Important
Archived historical snapshot — June 2026. This repository documents an earlier version of my Hermes setup. It is not actively maintained, and commands may be outdated. Use the official Hermes documentation for current installation and security guidance.
This is Leo Laporte's personal Hermes Agent configuration — an idiosyncratic, over-built, local-first AI agent setup that runs 24/7 on his own hardware. It's grown organically into something elaborate and very specific to how Leo works. It is not a minimal install. But every piece of it earns its keep, and together they turn a chatbot into something closer to a persistent external brain that's available from his terminal, browser, phone, watch, and speakers.
This repo exists so friends (and anyone curious) can see what's possible, then follow the included setup guide to get the foundation running and decide how far down the rabbit hole they want to go.
Hermes Agent is an open-source, self-hosted AI agent framework by Nous Research. It runs on your own machine (Linux, macOS, or Windows/WSL2), can use essentially any AI model, remembers what it learns across sessions, runs scheduled tasks while you're offline, and learns from experience by saving reusable "skills." It's the same category of tool as Claude Code or Codex — but provider-agnostic, multi-platform, and self-hosted.
- Host: Framework desktop (128 GB RAM, AMD AI+ 395), running CachyOS (Arch-based Linux)
- Philosophy: Local-first, self-hosted, privacy-preserving — Leo's data stays on Leo's machine
- Uptime: 24/7, UPS-backed, survives power outages and reboots unattended
- Agent name: Quicksilver (after Neal Stephenson's Baroque Cycle)
- Local-first main model running entirely on his hardware (no cloud round-trip)
- Fallback chain — transparently drops to a cloud model when the local server is down
- Per-task model routing — different models for delegation/judgment, cheap bulk work (compression, titles), and vision
- Automated model-strategy reviews — a weekly job benchmarks and reassigns the best model to each task slot
- OAuth over API keys wherever possible (uses existing subscriptions, no per-token spend surprises)
- Web UI (browser, the primary daily interface) — sessions, file browser, memory editor, cron UI
- Terminal CLI for power use
- Telegram bot — the main mobile/chat interface, with rich messages and streaming
- Discord and Slack gateways
- Home Assistant for smart-home control
A custom bridge (Obi-Wan Core) connects the agent to:
- Apple Watch & iPhone (voice messages, quick queries via Shortcuts)
- Rabbit R1
- ESP32 microcontrollers
- Sonos speakers — a voice server routes spoken replies to the appropriate room
Not one memory store — several, each for a different kind of knowledge:
- Durable facts & preferences that survive across sessions and model swaps
- Semantic long-term memory (Hindsight) for meaning-based recall
- Knowledge graph (Graphiti / FalkorDB) for entities and relationships
- Open Brain for captured thoughts, opinions, and observations
- Obsidian vault — daily notes, journals, coffee brew logs, exercise, travel
- Full-text search over every past conversation (never has to re-explain context)
Reusable procedures the agent saves when it solves something hard — coding, creative work, DevOps, research, note-taking, smart home, email, and more. Skills auto-load for matching tasks, so the agent gets better at Leo's specific work over time.
Dozens of cron jobs that run while he's away:
- Morning continuity briefing (calendar, mail, daily context)
- Mail digests (3×/day)
- Cost & usage dashboard
- Weekly memory consolidation
- Custom domain watches (e.g. a Peter Gabriel tour tracker, EV market watch)
- Local-model health checks
- Local text-to-speech (Kokoro, a natural voice) for spoken replies
- Push-to-talk voice mode and dictation via a local Whisper server
- Quiet-hours awareness — no voice output overnight
- Fastmail mail, calendar, and contacts via CalDAV + JMAP — the agent can search mail, draft replies, read his schedule, and look up contacts
- Self-hosted SearXNG metasearch (queries never leave his network)
- Local web scraper/extractor (Firecrawl-compatible, no cloud)
- Pulse — multi-source social search scored by real engagement (Reddit, HN, GitHub, YouTube, arXiv, and more)
- Parallel subagents for independent workstreams
- Integrations with Claude Code ("Kenobi" — the coding specialist), Codex, OpenCode, and OpenHands for heavy software work
- Bitwarden vault as the secrets store — keys materialized to tmpfs (RAM) at startup, never written plaintext to disk
- FIDO2 hardware keys (Token2.swiss) for login
- Local-first by design — he owns and controls all the data
- UPS-backed with clean shutdown on power loss and auto-restart when AC returns
- Treated as a server, not a workstation — survives reboots unattended
- Tailscale for secure remote access from anywhere
Start with the setup guide. It's written to be handed to Claude — paste the whole thing into a fresh Claude conversation and it'll walk you through the foundation interactively (detecting your OS, pausing at every decision that needs your input).
Honest scope note: the guide gets you the foundation — Hermes core, a model provider, the Web UI, memory basics, and an optional Telegram bot. Everything in the feature list above is what Leo has built on top of that foundation over time. It's a direction to grow in, not a one-shot install. The Hermes docs cover how to add each piece.
Works on macOS, Linux, and Windows (via WSL2).
| File | What it is |
|---|---|
hermes-setup-guide.md |
The Claude-driven setup guide (the main deliverable) |
README.md |
This file — overview of Leo's setup |
This is a personal setup, but corrections, suggestions, and improvements to the guide are welcome via issues or pull requests.
MIT — see LICENSE.
Hermes Agent is © Nous Research, MIT licensed. This setup guide is © Leo Laporte. Built with help from Quicksilver, Leo's own Hermes agent.