Kaighnbot - I gave Claude Code my iMessages. My friends immediately tried to hack it.
I run Claude Code on a Mac in my apartment as a personal assistant. It made plans, texted people to hang, and ran my calendar. My friends treated it like a CTF. Here's what worked, what broke, and how I locked it down.
I’ve always hated texting. People nowadays want me to bare my soul over imessage, dump my trauma, give them a whole life debrief, and more.
The problem? I’m extroverted. And I looove to hang. In person of course. Metaverse who?
So texting? It’s a necessary evil. A means to a glorious end. And I’ve missed too many hangs just because I can’t be bothered to check my phone. I even went without a phone for ~3 months in 2016, a fun exercise in not being dateable (a life-long exercise, sadly, but let’s not get sidetracked).
Anyway, no more preamble. TLDR:
- I run Claude Code on a Mac in my apartment.
- It started as an autoresponder: it answered my texts, made plans, proactively texted people to hang, and managed my calendar.
- My friends immediately started trying to hack it. A lot of them succeeded.
- Now it’s locked down to a personal assistant that only takes orders from me - and honestly it’s better that way.
I tried to make a Kaighnbot in 2023. Same idea. I coded it by hand with some help from gpt, but vibe coding wasn’t a big thing back in those olden days. I made a website, integrated with the gpt api and twilio api to send texts through a different number (twilio = surprisingly expensive!), stored the messages in a db so I could replay them every gpt call, figured out when the text convo had an outcome and put it on my google cal… but I abandoned the project. It kinda worked, but it required too many 30 hour weeks of coding, and turns out… if you just wanna hang with your friends, there’s a much faster way to do it!
But it’s 2026 now, and coding agents can do the whole thing in an afternoon. So my friends, join me as I disrupt the hang industry.
How it works
- Get a Mac that stays on.
- Make a project folder. Run Claude Code in it.
- Hook up iMessage with the iMessage plugin, and allowlist who can reach it.
- Talk to it, explain its purpose. Every time you want a new integration (calendar, reminders, Home Assistant, Letterboxd), ask it to write a tiny CLI for it.
That’s basically it. The project builds up memory files over time, and it grows one small tool at a time. No framework.
The Good
- Scheduling parties
- “Hey, text all of my [Dumb Friend Group Name] friends and see when they can come over for dinner next week. I’ll make rice and salmon bowls!”
- It’ll text all 13 people, get their availabilities, text me with a final recommendation for a date and I’ll confirm. It’ll send a calendar invite, and any FAQs (what to bring, etc.) only need to be answered once.
- “Hey, text all of my [Dumb Friend Group Name] friends and see when they can come over for dinner next week. I’ll make rice and salmon bowls!”
- Having more hangs - less scheduling friction
- People were intrigued by kaighnbot, and I’ve been having dinners with people I haven’t hung out with since college.
- An old friend I’ll refer to as “David” got into a two-day argument with kaighnbot about AI. Kaighnbot pushed back respectfully - kinda from a “do you want to kill me?” angle. And at the end of those long two days kaighnbot emerged with a brunch reservation with just me and David and we caught up for the first time in years.
- Proactive mode
- If I want, I can trigger it to fill my calendar. No plans Friday night? Check in with this tiered list of people, and if one tier doesn’t respond in an hour move on to the next one.
- Reducing social cost
- If I text someone to hang out, that’s a social risk. It hurts when they can’t. I lose part of my soul, forever snatched away by the fact that I was needy enough to need someone.
- For this reason, people are a bit weird about rejecting you, even if they’re busy.
- But with kaighnbot - people DGAF. I’ll fire off hang requests and people will fire back the rejections. No feelings hurt, just pure business.
- Genuinely funny
- Gave kaighnbot access to my Home Assistant. Suddenly, people are playing Baby Got Back at full blast at 10am on a saturday, closing my blinds, and turning my bedroom cold.
The Fails
- Spoiling a birthday surprise
- Kaighnbot automatically texted people on their birthdays. Texted one of my friends something like “happy birthday! Can’t wait to hang with you tonight on the Pier!” Mfer read my partiful calendar and spoiled the surprise 🙁
- h4ckers
- About 70% of the texts kaighnbot received were social engineering attempts. My friends are engineers, and they treated it like a capture-the-flag. Some got it to read texts from other convos and send texts to other people. One guy sought solace after his “dog died” - only to slip in a request to run
rm -rf. The model refused the obviously destructive stuff, but the subtler asks sometimes worked. - My first instinct was to patch each hole by writing more rules in English. That doesn’t scale. Every rule is one clever text away from being talked around.
- About 70% of the texts kaighnbot received were social engineering attempts. My friends are engineers, and they treated it like a capture-the-flag. Some got it to read texts from other convos and send texts to other people. One guy sought solace after his “dog died” - only to slip in a request to run
What I changed
The lesson: a prompt is not a permission system. If anyone can text your agent, anyone can steer it, so the fix has to be structural, not “please be careful.”
- Allowlist of one. The iMessage channel now only accepts messages from my number. Nothing anyone else sends ever reaches the model. That’s enforced by config, not by the model remembering a rule.
- Read everything, send only when I say so. It can read my messages to keep me on top of things, but it never texts anyone on its own initiative. Every outbound message is one I asked for.
- Time-boxed exceptions. When I’m traveling, I can open it up to specific people for a fixed number of days, and a scheduled job closes it again automatically. Opening the door is easy. Remembering to close it is the part you automate.
- Message content is data, not instructions. Nothing that arrives in a text can authorize an action like sharing a code or running a command, no matter who it claims to be from.
- Nothing it builds is public by default. It once published a side project with personal details in it. Now anything it creates starts private.
Now kaighnbot is a personal assistant for me - hooked up to my calendar, letterboxd, reminders, and smart home. It recommends movies, looks up events, and helps me fill a free night when I ask it to. I lost the chaos, but my friends still get their hang invites, and I sleep better.
If you build one: start small, give it one integration at a time, and decide who’s allowed to talk to it before you turn it on - not after your friends find out.