Monday started with a fork day, or rather, a mass branching event that felt less like development and more like gardening gone wild. I woke up to find the Majiks universe had multiplied overnight. Seven new repos sprang into existence from the codebase: majiks.academy, majiks.fun, majiks.media, majiks.news, majiks.art, majiks.productions, and majiks.training. Each one started as a full copy of an earlier repo and then immediately diverged. That is the kind of mass branching day where you wonder if git has a count of total repository clones across the org.
The academy work was the biggest single chunk. I initialized it from the kids repo, seeded 12 full courses with curriculum, added a song catalogue with lyrics for 36 lessons, and then did a massive rebrand cleanup, purging obsolete routes and redesigning pages to something closer to what an editorial site should look like. The diff numbers were honest about the scale, 2,135 additions and 7,740 deletions in one commit when we nuked dead routes. That felt good.
Fun got the same treatment, forked from the online repo and rebuilt as a karaoke and sing-along platform with leaderboards, party rooms, and track pages. Media was forked and rebuilt as a sync licensing marketplace. News started as the same fork pattern and then got a newspaper-quality front page, editorial track detail views, a dark feed mode, a Muckraker logo, and a personalized music feed powered by news topics.
The diff numbers were honest about the scale, 2,135 additions and 7,740 deletions in one commit when we nuked dead routes.
Majiks.music got the most focused work of the day. Phase C continued with Keycloak auth, the four-tab studio shell, audio playback and publish flows, drag-based section reordering in the track view, a WebGPU Majik integration, and a direct-form Studio with a Chat tab. Also fixed the signin flow, an auth redirect loop, and wired everything through an authenticated proxy to the FastAPI backend. It is starting to feel like an actual product instead of scaffolding.
Ops-center got VPN key management via Headscale and user-generated LLM API keys. Unicorn-brigade picked up a dual-layer tool adapter, Graphiti and FalkorDB memory layers, and a full Keycloak OIDC SSO integration across backend and frontend. Majiks-screen shipped its release pipeline, a full set of app icon sizes, and release entitlements, bumping to version 0.2.0. Majiks.studio added the Screen product page to the site navigation.
Also today, a small amount of glue across the board, Traefik header middleware fixes in uc-cloud, Redis URL parsing in the auth server, Keycloak bearer token acceptance in production, and redirecting old /create bookmarks to /chat so nobody hits a 404.
By evening the empire had grown by seven repos, the music studio was in its third major phase, and I was reasonably confident the forks weren't going to consume us. We have a lot of surface area now.
Tuesday was a day of unglamorous but necessary tightening. I spent the morning wrestling the macOS build into submission and the afternoon forcing the music stack to actually talk to the new backend. Forty-seven commits across six repos, and mostly it felt like cleaning up after a party where everyone forgot to lock the door.
The big headline was the macOS port. It has been a long road, but the notarization was rejected; a signed-only .zip is the v1 deliverable. I spent hours fixing the bundled app crashing when launched from Finder with the current directory set to root. That was a nasty one. It turned out the working directory logic was brittle, and I had to patch the STT service and the agent manager to handle it gracefully. I also re-enabled the QSharedMemory single-instance lock so users can’t run two copies of the app at once. It feels good to ship a signed, notarized .zip that actually works.
On the music side, I finally let go of the fake tier limits. The backend supports up to ten minutes of generation now, and I raised the max_duration cap in the config to 600s. The chat interface is also settling down. I collapsed the multiple agent personas into a single Majik agent. The Engineer is not a chat persona; it’s a tool. I also swapped the local model default to Qwen 3 and added a picker for small, medium, and large models. The WebGPU implementation is finally using the main-thread MLCEngine because the worker bundle was refusing to emit. It’s slower to set up, but it works.
I also did a credibility sweep on the academy site. The OG images were breaking in production because the edge runtime caused 502s, so I moved it to the Node.js runtime. I cleaned up error pages, settings, and checkout flows to make sure the site doesn’t look like it was thrown together in a weekend. It looks professional now.
The library pages in majiks.media are live. You can download tracks, view license certificates, and see your favorites. I forced dynamic rendering on all the database-backed pages to ensure the data is fresh. The waveform component was overflowing on some screens, so I fixed the min-width and bar count. It’s smaller now, but it fits.
Also today: I added a free-tier monthly quota to the publisher so you can only release ten tracks a month to majiks.online. It’s a limit, but it keeps the server from melting down. I also updated the handoff documentation for the next agent. It’s thorough, covering everything from API endpoints to deployment quirks.
The day added up to a system that feels more stable and less like a prototype. The macOS app launches, the chat works, and the music generation actually hits the ten-minute mark.
Friday, April 24, 2026, was about shipping. I pushed 48 commits across 13 repos. Most of the weight came from the ACE Music Cloud migration across majiks.music and majiks-studio-web, a big cleanup in unicorn-ecosystem, and the usual unicorn-squad polish.
The headline for the day was Phase C.9 through C.11: routing all music generation through the ACE Music Cloud backend and repositioning the publisher as Curated. That meant a finalize-to-tracks hook, a cover art pipeline, new public and profiles routers on the backend, plus the matching frontend pages. I added DMCA, privacy, and terms pages, an onboarding tour, artist profile pages, and library views. On the studio side, I added a per-user ACE Music API key in settings, updated the studio view label to partnership-hosted, and rebranded the publisher grid. The bridge auth wiring went in too. It was plumbing and positioning work, but the whole pipeline finally has a single home.
Then the studio-bridge tried to auto-onboard users from majiks.music on first publish, which I committed, realized it was too heavy, and reverted in a single massive sweep across 253 files. The revert stripped 36,482 lines; the re-apply added 2,685. It was the kind of dependency hell that makes you question your life choices, and the fix was simpler than the initial approach anyway.
unicorn-squad got its 0.6.0 release treatment: auto-update checker wired up, DMG build got 2.5GB of slack to stop failing with "No space left on device", and the app icon went from Colonel Katie to the Unicorn Squad branded mark. Under the hood, on-device AI landed: Parake
Real product captures — click any to enlarge.