We're building Codex Alliance in the open.
Not a launch party.
A quiet place for people who make things —
and still want somewhere to belong.
Build together. Belong anywhere.
This account is our diary of that work.
The app is still in internal testing. It hasn't gone out to outside testers yet, and these failures only showed up on my own phones. I'm hoping to launch it officially this week, and I'd love for you to come try it when it's out. Early on, I'll be giving away 100,000 free Founding Member spots. On one hand, this is a real developer community. On the other, I hope that through everyone's experience and guidance, we can make this app better together. There's quite a lot to play with in it.
I'm a solo founder. It's Monday morning. My app's DMs are broken, and honestly I don't know how to fix them.
So I'm just going to tell you everything.
Codex Alliance runs on Apple's CloudKit. No third-party servers. I picked it because I'm one person and I can't babysit servers at 3am. Apple holds the data, your phone keeps a local copy. It sounded so simple.
Last night I sat with two iPhones on my desk. Two accounts. I typed "hi" on one and stared at the other.
The sender saw its own bubble after about 3 seconds. The other phone? Nothing. I kept staring. 30 seconds. 60. 110 seconds. Still nothing. I tried the other direction. Timed out too.
The worst part: one round earlier that same night worked fine. So it's not even broken in a way I can point at. It just... sometimes doesn't arrive.
Photos are worse. I can't even give you a "slow" number, because I never got one through. The sending side failed. The other phone never showed it. The next try stopped before it even sent. It's past midnight and I'm watching a loading spinner like it owes me money.
This morning it got bad enough that I just told Codex: drop everything, DMs only, real phone only, check the app and the CloudKit side together. That's where we are right now.
Some of the slowness is my own old design. Publishing one post used to take a chain of writes plus a check back from the cloud. If one step stalled, it sat on "not confirmed" forever. The Creator Center waited up to 12 seconds for the cloud before showing you anything. Twelve seconds of a blank screen, on your own posts. The fix is to show what's already on your phone first. It's written. It's just not in the test build yet.
Because this morning we also had to roll back. Overnight we shipped a new sign-in screen and a test build that fixed a dead search button and made Arabic read right-to-left. Then to keep the new sign-in, we went back to an older base, so some of those fixes have to go back in by hand. One step forward, half a step back.
Small wins, so this isn't all crying: in the simulator a DM bubble now shows up in 12 to 28 ms. On a real iPhone 13 the app opens in about 2.7 seconds. And the website got some love: our little cloud is back in the middle, and it now speaks 131 languages, up from 50.
But fixed in code is not fixed in your hand. Until two real phones can talk to each other, I'm not calling anything done.
So, a real ask, not a rhetorical one: if you've ever built chat on CloudKit, how did you make messages between two devices feel instant? I'll read every reply. I need it.
#Codex #BuildInPublic #CodexAlliance
The app is still in internal testing. It hasn't gone out to outside testers yet, and these failures only showed up on my own phones. I'm hoping to launch it officially this week, and I'd love for you to come try it when it's out. Early on, I'll be giving away 100,000 free Founding Member spots. On one hand, this is a real developer community. On the other, I hope that through everyone's experience and guidance, we can make this app better together. There's quite a lot to play with in it.
Actually, nothing is broken on either side. The only problem is that sending is way too slow. Plain text takes about 3 seconds to land in the other person's chat. A photo takes around 10 minutes before the other side gets it. Here's how our flow works: when A sends a message to B, A's message is encrypted first, then synced to CloudKit. B's side then automatically matches the key, and only then can B receive the message and unlock it in the chat. It's just too slow. I'm trying to figure out how to speed this up, because syncing to CloudKit takes time, and decrypting also has to wait on CloudKit.
We use both, and it still broke last night. All the data is in CloudKit's public database. DMs are encrypted on the phone first and then saved, and photos and text are packed into the same encrypted package. So the other side can't show a message the moment it sees the record. It has to download the whole encrypted package, check the sender's profile and the conversation, and then decrypt it before the bubble shows up. The silent push is only used to wake the app up. While a chat is open, it checks every 4 seconds, and it fetches again when the app comes back to the foreground. Last night I tested on two real phones: the sender saw its own bubble in about 3 seconds, but the other phone still hadn't received it after 110 seconds. When I sent a photo, it took 5 minutes for the image to arrive on the other side. I even laughed and said to Codex: if every single message in a chat means waiting 2 minutes for the other person, who would keep waiting? I chose CloudKit because I wanted it to be clean. Users' rights and privacy are stored on their own devices, 1:1 chats are encrypted between the two people, and the backend can't see what's in a DM. It only stores the encrypted text. But the lag and the slowness are something I really can't stand.
This morning I took the waiting line out of the Codex Alliance app. By lunch I had to build the room behind it.
Here's the follow-up. Until now, every post and every Codex project you shared sat in a queue until I approved it. It felt safe, but slow, like you needed permission to talk. So we flipped it. Now what you share goes live right away. Reporting, blocking and the automatic pause on flagged posts all stay.
But the moment you stop checking things before they go up, you need a real plan for after. So today Codex and I built a small control room on my Mac, the place where I handle reports, support requests and mutes when something goes wrong.
First rule: the control room checks who I am before it opens, and again before every serious action. If the account doesn't match, it locks itself. Every action leaves a written record, and if saving that record fails, the app keeps it and tries again later instead of quietly losing it. Only I will ever see this screen. That's exactly why it has to be honest.
Then came the hard question: what happens when a shared project has to come down? Codex wrote me a one-page proposal and asked three questions. I said yes to all three. Taking something down means hiding it, not deleting it, so a mistake can be undone. Only my checked account can do it, with a second confirmation every time. And the creator sees a clear "hidden" label with a way to appeal, instead of their work just vanishing.
The third answer was the uncomfortable one. If someone already saved a copy on their phone, it can stay there until that phone refreshes. No app can take back what's already been copied. I'd rather tell you that than pretend.
Where this honestly stands: the control room is finished on my machine and passes all its checks, but none of it is in anyone's hands yet. The takedown rules are approved, not built.
The rest of the day was small stuff, and I loved it. In the Codex area, the Following and For You switch sat at the top of the screen. My AI teammate argued it belonged there, like most apps do. I pointed out that every other part of our app keeps that switch at the bottom. It agreed it had missed that. It's next on Codex's list, because one app should feel like one app.
I also caught a bug while chatting: typing on a Chinese keyboard, the gray hint text stayed on screen and overlapped the letters I was still typing. I've asked Codex to fix it everywhere that kind of box appears. And I asked it to rename a settings line from "Storage" to "On This Phone", because people should know exactly where their stuff lives.
Where this is going: open by default, careful when it matters, and honest about the limits. That's the kind of community I'd want to join.
So I'm curious:
if you run a community, do you check posts first, or let them go live and deal with problems after?
And if your post ever got taken down, what would you want to see?
We're building the Codex Alliance app and community.
#CodexAlliance #Codex #BuildInPublic
@hi66294029 Ha, exactly. That line is a promise to myself. Once the rebuild lands on a real iPhone, I'll post the before-and-after numbers: cold start, time to post, and time to open a chat. No vibes, just stopwatch ⏱️
A week ago I thought our biggest problem was a few buggy screens.
Today I learned it was something deeper: where the app keeps your stuff.
Here's what happened. When we built the Codex Alliance app, we put a lot of each person's data in their own iCloud. It sounded neat. Your data, your account. But in practice it made the app heavier than it needed to be. Every tap had more places to check, more ways to get stuck, and it quietly ate into people's personal storage.
So today we sat down with Codex and didn't write a single new feature. We wrote down the truth first.
We added a way to see exactly why the app fails to connect: is it your account, your network, or us? Then we went through every piece of data, file by file, and listed what still lives in personal iCloud. Boring work. But you can't fix what you won't look at.
The new direction is simple: the small personal stuff, like your bookmarks, mutes, what you've already read, your settings, lives right on your phone. The shared stuff, like posts, comments, profiles and messages, lives in one public place for the app. Fewer moving parts. Faster taps. And nobody's existing data gets deleted along the way.
The hard part is group chats. Right now a group lives in the owner's iCloud, which means one person is carrying everyone. We're going to rebuild it the way our private messages already work: in the shared place, but locked so only members can read it. That's not a one-day job, and I'd rather be honest about that than pretend.
A few things did get finished today. Creator Center now looks like a normal phone settings page instead of a custom maze. And we're adding live name checks: type a username, pause half a second, and you'll know right away if it's taken. It even catches lookalike tricks like extra spaces. Because nothing feels worse than setting up an account and then finding out your name doesn't work.
We're also cleaning out old code nobody uses, and we'll measure app size, startup time and memory before and after, so "faster" is a number, not a vibe.
Where this goes: once the foundation is solid, the app should feel like it belongs on your iPhone. Reply to a message right from the notification. Share a photo into the app straight from Photos. See upload progress on your lock screen. Find saved posts from iPhone search.
The lesson for me today: the features people see sit on top of decisions nobody sees. Get the bottom layer wrong and every new feature makes it worse.
If you're building something too, I'm curious:
have you ever had to stop adding features and rebuild the foundation first? Was it worth it?
And if you'd use one of those iPhone features first, which one?
We're building the Codex Alliance app and community.
#CodexAlliance #Codex #BuildInPublic
Fair point, and thanks for flagging it early. Analytics and support are exactly where the public DB gets awkward. We're staying Apple-only, so our own macOS operations console reads the public DB directly, and we're building out its query and support tools now. Basic metrics come from App Store Connect. If we outgrow that, a nightly export goes to a machine we own, not a third-party service.
Thank you for taking the time to write this out. It's genuinely the kind of feedback we need, and honestly it points at something we hadn't written down yet, so it's a place where we need to improve.
You're right that derived data is the one that slips through. Codex Alliance has it too: post counters, heat stats, search. It isn't personal and it isn't public content, so without a rule it really would land wherever was convenient that sprint.
We're adopting your third line. Since we run no servers of our own, "a service we control" for us means our own CloudKit public database. Derived data is computed there from public records, or on the device from local data, and never leaves those two places. We're also taking your second clause: every new index or counter must name the field it derives from. You're right that it's the clause that catches the quiet copy someone makes for a dashboard.
You called it an addition rather than a correction, but to us it's both, and we're grateful for it. This is what we started Codex Alliance for: builders sharing, learning and getting better together, exactly like this conversation.
@RAHULKESHR31818 That's exactly why we dropped it. No own backend, though. Shared data lives in Apple's CloudKit public database, and personal data stays on the phone. Fewer moving parts, no hidden sync conflicts.
@jacksunny18 Agreed, and that's exactly what we did. We wrote the list down: on-device for personal data, CloudKit public DB for anything other people need to see. Anything new has to fit one of those two, or it doesn't ship.
@aksenHQ Personal stuff (bookmarks, mutes, read state, settings) now lives on the device. Shared stuff (posts, comments, profiles, messages) lives in Apple's CloudKit public database. No per-user iCloud sync anymore.
@thsottiaux that "won't rest" energy hits. from the community side we're the same — language screens, keyboard bugs, publish path — keep pushing Codex Alliance until it feels right on a real phone
#CodexAlliance
Confirmed so far: the text-post entry regression passed; first tap on the comment box, continuous typing, and typing again after collapsing the tools all work. One test failure was only a different cursor position after tapping the input — I changed it to check that the full text is still there. Real-device automation just hit one init timeout and never reached the business steps; retrying now, and that can’t count as a pipeline pass.
Real device is installed and verified as build 51. Running focused tests on comments, the side drawer, and the system video picker. The image-swap data fix just added will go into final build 52; the official distribution package is being generated, then it’ll be installed via TestFlight and retested.
Also found a concrete bug in the image pipeline: when swapping to another image, the code only compared counts, so it could keep using the old image data; and if the new image hadn’t finished loading yet, it could wrongly reuse data from the old draft. I’ve fixed both places and added a check for stale read results so a fast image swap can’t let an old task overwrite the new selection. The final package will include this fix.
Also found a concrete bug in the image pipeline: when swapping to another image, the code only compared counts, so it could keep using the old image data; and if the new image hadn’t finished loading yet, it could wrongly reuse data from the old draft. I’ve fixed both places and added a check for stale read results so a fast image swap can’t let an old task overwrite the new selection. The final package will include this fix.