Auryn
Auryn is the personal assistant I built for myself. It runs on a computer at home, handles email, calendar, reminders and the lights, and writes a lot of my code. This page is about the parts that took the longest to get right: what it's allowed to do on its own, and when it asks me.
Numbers are as of October 31, 2026. Every picture on this page is a recreation, not a screenshot.
What it replaced
-
Approval textsA card on my phone
In October, 22 texts asked me to approve 15 changes, and 7 of them were repeats. Now each decision is one card that I answer with a swipe.
-
An approval queueAct and log it, or refuse and tell me
Over five days in September, 65 requests sat waiting for me and the household work stalled. Now an independent reviewer approves anything that can be undone and logs it. Anything risky is refused, and I get one line saying so.
-
The Software FactoryShort briefs and a second opinion
The factory built whole apps from a long spec. It needed about a line of spec per line of code, and its second version never ran a job. I retired it in October.
-
Remembering what runs whereA live map that checks itself
Services, devices and vendors are on one map, linked by what depends on what, with fix-it notes attached to the ones they cover.
-
My old dictation appWordhold, running at home
Speech is turned into text and cleaned up on a computer in the house. It's the only dictation I use on my phones and laptop.
By the numbers
- 195 things on the map: services, devices, vendors, with 473 links between them
- ~130 tools Auryn can use
- 463 build jobs handed to coding agents in October
- 49 of the first 54 Deck cards answered in six days, a median of 9.6 minutes after they arrived
None of the 54 cards had to fall back to Auryn's default. One household is the whole user base.
How it's put together
A map of the house that checks itself
Auryn keeps a live map of everything it looks after: services, devices and vendors, linked by what depends on what. Seventeen of the key pieces have step-by-step fix-it notes attached, so for those the first step is already written down.
Anything with a health check is probed every five minutes. Five sync jobs, running from every 15 minutes to once a night, keep the rest matched to reality. When I audited the map in August, the parts filled in by syncing had no drift at all, and the parts I'd typed in by hand had all of it. Now everything that can sync, does.
Small jobs with a brief and a second opinion
When something needs building, Auryn writes a brief and hands it to a coding agent working on its own copy of the code. Every brief starts with a job card: which requirement the job serves, what's out of scope, and how big it's allowed to get.
Before anything merges, the agent has to show proof it works, and a second model reviews the change against the brief. A finding gets fixed, argued with evidence, or declined and written down. A timeout or an error doesn't count as approval.
Wordhold: turn on overnight learning
- Serves
- R14, overnight learning
- Out of scope
- Anything that sends audio or text out of the house
- Appetite
- Medium
- Brief
- Build
- Proof
- Second opinion
- Merged
The Deck: only real decisions reach me
Some questions Auryn can check for itself, like whether code is safe to merge. Those it decides, and it makes the result easy to undo. Others only I can answer: does this look right, should we spend this, can we message this person. Those become cards.
- One decision per card, and one card per decision.
- The card shows the thing itself, Auryn's own pick, and why.
- It says what each answer will do before I give it.
- If I don't answer in 48 hours, a card that can be undone gets Auryn's pick. Money, people, deleting things and access always wait for me.
Approve
Swipe fixed: does it feel right now?
My pick: good. The card now tracks your finger one to one, and up or down can't be flicked by accident.
← badgood →
- Right good
- Left bad
- Up add something
- Down rethink
What didn't work
The Software Factory. It turned a written spec and a list of test scenarios into a working app, then kept fixing the code until the scenarios passed. Its best run rebuilt cal.com, the scheduling app, and passed 99% of 127 scenarios. But the spec had to be about as long as the code, and the second version never ran a single job. The lesson I kept: write the checks from the requirements, keep jobs small, and let me approve the checks rather than the code.
Failures that looked healthy. Four separate messaging faults all reported success from the inside. Now every write is confirmed by reading the result back, and “queued,” “sent” and “verified” are tracked as different things.
Asking too often. The approval queue was meant to make Auryn safe. In practice every request it raised dead-ended waiting on me, and the household work stalled. Refusing the risky things and telling me in one line works better than a queue nobody gets to.
Limits
Auryn serves one household, which isn't scale, and it isn't for sale. Personal details, addresses and anything on my home network are left off this page on purpose.
Want one chore handled like this?
I set up the same kind of help for small businesses, one recurring chore at a time, on accounts you own.
See how it works