Hermes is an open-source AI agent that lives on my own machine, answers in Telegram, runs shell commands, drives a browser, and — most importantly — remembers. These are the projects and automations it operates for me: a household operations stack built in about five weeks of conversation.
Three properties make the difference between a chatbot and an operator.
It executes on my desktop with terminal, filesystem, and browser access. Real files, real commands, real deployments — nothing uploaded to a third-party workspace.
A Telegram DM is the control room. I text it from my phone the way I'd text a chief of staff; it replies with results, files, and voice notes.
Everything it learns becomes a written skill — a procedure with pitfalls and verified commands — or persistent memory. Each task makes the next one cheaper.
Full products, not demos. Each one started as a sentence in a chat window.
A recipe library with family ratings feeds a weekly meal-picker web app: pick dinners from real recipes we've actually cooked, get a merged, quantity-collapsed shopping list, and store the plan for next year's rotation. Recipes live as markdown, the picker deploys to a public URL in one command.
Ratings are one-to-three stars recorded per recipe after dinner, so the picker learns what the kids will actually eat — not what looks good on paper.
A private, one-listener podcast: every night it reads the long-form articles I've saved, writes a magazine-style spoken script, and generates an episode with an AI voice. A YouTube channel I follow gets ingested the same way. Served from my machine over a private VPN, straight into Apple Podcasts.
Source articles are kept verbatim for attribution; audio is transcoded to broadcast-safe MP3; the feed verifies with byte-range requests like a real CDN.
A custom home screen on an e-ink panel that redraws itself with fresh renders: tomorrow's weather with clothing advice, bitcoin price beside its 20/50/200-day moving averages, moon phase, family birthdays, and the school lunch menu.
HTML screens are designed, screenshot-verified, and pushed through the device's API on a schedule — the layout code is mine, checked pixel-by-pixel.
An interactive explainer for the US individual tax return: click any line and see the computation traced backward through schedules and source documents. Includes a cross-check engine that reconciles every stored figure, and a specification for a full rebuild written by the agent itself.
Working prototype with sample data; a detailed spec exists for the production version with a real calculation engine.
The daycare logs our son's diapers, bottles and naps in one app; at home we use another with no API and no import. So an Android emulator runs the real app as one more device in our sync group, and a script types each entry in through its own forms, reading every one back before moving on.
Won't write unless the app shows the right child, stops at the first mismatch, and halts itself after three failures. Read the write-up →
Scheduled jobs that fire without me. When one breaks, the agent diagnoses and fixes it from the run logs — usually before I've noticed.
Smaller builds and utilities, all live.
A print-ready magazine compiled from saved articles — layout, proofreading gates, and single-copy print ordering.
A local budget application with bank aggregation. Financial data never leaves the house.
Household federal and state tax estimators, with data flows kept strictly private.
An Obsidian knowledge base with a compiler and query layer for planning across sessions.
A photo-metadata CLI that extracts timestamps and GPS into a YAML feed for a family location map.
An offline-first children's operating environment — source-available, carefully licensed.
Declarative capture and restore of the whole desktop — used for a complete machine-to-machine migration.
A keyboard-driven terminal application for market data, styled and extended pair-programming style.
Web articles converted to files for a small e-ink reader, with typographic tuning per device.
Word, Excel, PowerPoint, and PDF workflows — reading, editing, filling, OCR — plus meeting notes turned into tracked action items.
Terminal email triage over IMAP/SMTP, and full GitHub workflows: issues, pull requests, code review, CI status.
A static site served from my own machine through an outbound tunnel — no open ports on the router.
A private CalDAV calendar the agent can add events to, reachable from my phone over a VPN.
Hue lighting control, direct-to-printer IPP printing, Linux troubleshooting, audio denoising, Bluetooth pairing, maps and routing.
The compounding layer: written procedures the agent loads before it works on something, so lessons are never paid for twice.
Each skill is a small operating manual — when to use it, exact commands, verified workflows, and a pitfalls section written from real failures. When a task goes wrong, the post-mortem becomes a skill patch. The library now covers everything from deploying web apps to transcoding audio for finicky podcast clients, to driving an Android emulator over adb.
Personal skills are layered on top of a shared open-source baseline, and the same library runs across a terminal CLI, a desktop app, and this Telegram gateway — one brain, many surfaces.