Things friends always ask.

Roughly 6 months of evenings + weekends to get to what's on the operating map today. The first usable version (one scheduled job + Telegram approvals) was about a weekend. Lyra (then called Sarah) took maybe three weeks. The Obsidian wiki layer was a few weeks of refining. Most of the rest was incremental.

Don't budget six months for yourself. Budget one weekend for the first useful version, then add a workflow per week as you notice patterns.

No, but you need to be comfortable copying code into a file, running shell commands, and reading error messages. If you've ever set up a scheduled job or written a Python script, you're qualified.

If you've never opened a terminal: spend an hour learning the basics first. Then do the 30-minute quickstart.

Linux works the same way (LaunchAgents become systemd units, but the pattern is identical). Windows is harder — you'd want WSL2 + native cron alternatives, and a lot of the Hermes-specific tooling assumes Unix.

For a Windows-first experience, look at cloud-based alternatives like n8n on a small VPS, or just run the AI parts locally and the scheduling parts on a Linux droplet for $5/month.

Yes. The pattern doesn't care which model you use. Substitute the OpenAI API where you see Anthropic. Switch the model name from claude-haiku-4-5 to gpt-5-mini. Adjust the request body for OpenAI's slightly different format.

That said: the reason I use both Claude and Codex (OpenAI) is because they catch each other's mistakes. Two AI vendors > one. Pick whichever for night one; consider adding the other later.

Three reasons: bots are dead-simple to set up (no admin approval, no app review), the API is excellent, and topic-channels let me organize nine agent lanes inside one app instead of nine different inboxes.

Slack works for team setups. iMessage doesn't have a usable bot API. Discord is fine but feels overbuilt for a 1-person operations channel.

This is the single most important question, and the answer is: the system is built so a hallucination can't actually do damage. Risky tools (account or money questions, legal issues, posting work orders, moving money) all stop at a Telegram approval gate, and scheduled jobs can't run risky commands at all. Routine, fact-only replies (property facts, prices, how-tos) can go out automatically, but only behind safety filters that hold anything uncertain for me.

The hallucinations that DO sneak through are the small ones: a wrong date in a summary, a misclassified intent. Those I catch on the morning review. They're embarrassing, not dangerous.

launchd restarts the gateway; a Hermes job re-fires missed scheduled runs every 20 minutes; ops monitors check credentials and the browser profile daily and alert the ops lane.

Net result: I've had maybe one outage I noticed in 6 months, and the system told me about it before I noticed it organically.

Honest answer: it goes to whatever vendors I'm calling. Anthropic, OpenAI, xAI, ElevenLabs, Twilio, Cloudflare, Google. Each one is a separate trust decision with separate ToS.

What I keep off the page on this site: tokens, chat IDs, tenant data, addresses, financial info. What I keep off vendor APIs: anything I wouldn't want public. There's a public-safety boundary I respect on every workflow.

If your data is regulated (HIPAA, attorney-client, etc.), this architecture is wrong for you. Use on-prem inference instead.

Emergency-class events fire on a fast-path: dedicated Telegram channel + SMS to my personal number, in parallel. If I don't notice both, I have bigger problems than the AI system.

For non-emergencies, the answer is: nothing terrible happens. The work sits in a queue. I review it next time I check the channel. Only routine, fact-only replies go out on their own; anything that needs judgment waits for me.

That's actually the whole point of the architecture. Every scheduled job can be disabled with one line. Every LaunchAgent can be unloaded. The Obsidian vault is just a folder of markdown files — they don't go away when you uninstall anything.

The reverse migration looks like: launchctl unload ~/Library/LaunchAgents/*.plist, comment out your crontab, cancel your subscriptions. The data and the notes stay yours.

For me, ~1-2 hours/week. Some weeks zero, some weeks four when something breaks or I'm tuning a new workflow. Average is ~90 minutes.

Most of that is reading the morning Telegram thread, catching anything weird, and occasionally tightening a prompt. Not “fixing the system” — more like “operating the system.”

The 30-minute quickstart costs about $20/month total. That's a real, useful loop — daily AI brief, Telegram approval, one scheduled job. You can stop there and most people probably should.

The full setup (its AI model plans alone run about $400/month) makes sense if you're running a real business with recurring operations. If you're an individual, the cheap version is probably enough.

For me running a property business: it's saved money. The alternative would have been hiring a part-time admin (~$2,000/month at minimum) or accepting more dropped balls. A few hundred dollars a month for a system that handles intake 24/7 is a clear win.

If I were running it as a personal productivity setup with no business behind it, I'd probably break even on the cheap version and lose money on the full version. Be honest about the math for your situation.

Both are great for “API A → API B” pipelines. Where they fall down is anywhere you need real reasoning — classification, drafting, judgment, multi-step decisions. They also charge per-task at scale, which gets expensive fast.

The sweet spot for the Hermes-style pattern is where you need judgment in the middle of the pipeline, not just data movement. If your workflow is “when X happens, do Y” with no thinking required, Zapier is probably easier.

Closer fit. n8n is the open-source Zapier and supports calling LLMs as nodes. It's a legitimate alternative if you prefer a visual workflow builder over writing scripts.

What you give up: the “everything is plain markdown + cron + a folder of scripts” simplicity. n8n adds a database and a UI you have to maintain. Pick based on whether you'd rather edit JSON nodes in a UI or edit Python scripts in a text editor.

That's actually a great starting point and I still do it daily for ad-hoc questions. The Hermes-style system isn't a replacement for chat — it's a way to make AI work happen without you starting it every time.

If your workflow is “I think to ask Claude something whenever I remember,” chat is fine. If your workflow is “this needs to happen every Tuesday at 9 AM whether I remember or not,” you need the cron + approval pattern.

Honestly? For some workflows, a VA is cheaper and more reliable. The system shines when the work is high-frequency, low-judgment (Lyra's call intake), or needs context the VA wouldn't have (the wiki + memory layer).

Best case is probably both: a VA for the things that need a person + an AI system for the patterns that benefit from machine speed. Don't think of it as either/or.