Blu_sundae

In my world, everyone now builds internal tools with AI agents: an analyst's dashboard, a finance tracker, a nightly report. The code takes minutes. Then it stalls. Where does it run? What will it cost? Who can open it? What stops a bad deploy or a runaway bill? So these apps either die on a laptop or land on an always-on server that costs $50 a month to run a job that takes two seconds. ShipCube is the missing step after the code. Your agent (Claude Code, Cursor or Muse Code) connects over MCP and asks ShipCube to ship the app. ShipCube runs it in a locked-down sandbox with no network, measures it, and a classifier picks where it should run. Meta's Muse Spark then explains that choice in plain words, compares the alternatives with their prices, and recommends the app's caps. Muse never sees the code, only the measurements. The app deploys to Google Cloud as a private URL behind identity-aware proxy, with hard spend and rate caps. A person approves the risky steps: first deploy, cap raises, sharing with a whole domain, deletes. The agent can't approve its own requests. Apps get Muse through ShipCube's AI gateway with per-app token budgets. For data, they query named measures through shipcube-data, our governed semantic layer, instead of raw SQL: wrong column names are refused with a hint, and a typical query reads only the partitions it needs (6 of 20 in production). When the data model needs to change, Muse drafts the migration and a person approves it. Result: a nightly job costs $0.10 a month instead of $49.93, and a heavy forecast gets 2 GB only while it runs ($2.79 vs $57.83). Built 27 Oct – 1 Nov 2026. Both repos run fully on a laptop, no cloud account or key needed: - Platform: https://github.com/Kunal-Shahi/shipcube - Data layer: https://github.com/Kunal-Shahi/shipcube-data