Hey - small Aussie team here. It's 1am and we just open sourced ContextUI.
It's a local-first desktop app for building and running visual AI workflows. TSX-based UI modules, full filesystem/CLI/GPU access, works offline with no cloud required.
Workflows can be published to the Exchange - others install and run them locally with deps resolved automatically.
Agents can create workflows and immediately test/run them in the same runtime.
Desktop app is Apache 2.0. ~300 users, ~230 workflows published.
Ubuntu. Its great. So much cleaner and userfrienly then Microsoft. Definitely don't need to be a dev to work in Ubuntu anymore. Honestly, I don't know how microsoft is holding its base.
cool idea and nice execution. highlights how far browser-native tools have come. curious how deep the DSP actually goes though — is this more a creative sketchpad or something you could realistically build full tracks with long term.
Thank you! The idea is to build full tracks, hence in the app you will find a timeline and a minimap, and there are builtin tools like labels/timeline functions to make arrangements over time. The DSP is state-of-the-art, i'm doing codegen of all the permutations of scalar/audio parameter inputs so the execution is always optimal. There is still room for improvement of course.
good breakdown of the tradeoffs but it kinda stops at critique. explains why age verification becomes invasive but doesnt really propose what a workable alternative looks like. feels like we need concrete models or architectures not just pointing out the trap.
local-first tools make sense and new patterns/protocols will emerge from that. but convenience always wins, so web utilities won’t disappear — they’ll just evolve.
Most of them will have much fewer visitors in the future, so their ad revenue will decrease to the point that they may not even consider keeping their domains. People will have their own toolboxes that they build themselves. You could say these tools evolving into something more personal and customized.
feels like more solo builders are realising agents arent about intelligence, they’re about constraints — DSLs, deterministic layers, boring hybrid architectures.
you stop writing code and start designing guardrails. curious how this scales though.
Thanks JB_5000 — really appreciate you putting it that way. You're spot on: the whole point is constraints over intelligence. Guardrails (DSL, schema, deterministic replay, boring hybrid) are what actually make it production-usable.
On scaling: so far it's handling ~3k trial users + growing paid base with low four-digit RMB yearly infra (queue-driven scale-to-zero, Redis cache, R2 for artifacts). The real bottleneck is still alignment quality (good artifacts + human gates), not the constraint overhead itself. Haven't hit hard walls yet, but I'm sure 10x–100x load will expose new ones.
How about you? Have you seen constrained agents / deterministic layers scale well (or break) at larger sizes? Any guardrails that worked surprisingly well for you?
Interesting benchmark, but worth noting the methodology: skills are generated before the task, with no feedback loop. In practice, useful skills tend to emerge from doing — you attempt, observe what failed, then codify what worked. Generate → execute → observe → refine. The paper tests cold generation, which is a different (and less realistic) setup.
It's a local-first desktop app for building and running visual AI workflows. TSX-based UI modules, full filesystem/CLI/GPU access, works offline with no cloud required.
Workflows can be published to the Exchange - others install and run them locally with deps resolved automatically.
Agents can create workflows and immediately test/run them in the same runtime.
Desktop app is Apache 2.0. ~300 users, ~230 workflows published.
Tell us what's awesome and what's terrible