RocketRide × SCU Buildathon · 2026
Leverage
The intelligence resource manager under the model you already pay for. Cheaper models do the work, a checkpoint survives every failure, and the repository's own tests decide what counts.
The problem
Shipping with a frontier model in the loop means paying frontier prices for boilerplate, and losing the whole thread when a rate limit lands. The expensive model is asked to do everything, including the fortieth test, and when it fails mid-task the next attempt starts from zero.
How it works
Leverage sits under the model you already pay for and turns a repository plus a goal into a task graph, each task held to the tests that define it. An auction scores every model it can reach, local runtime, free routes or a subscription seat, on task fit, verified history and cost, then hires the winner for that one task.
When a worker hits a limit or fails a test, Leverage writes down what it understood, decisions, files touched, checks already passing, what is left, and hands that checkpoint to a replacement. Nothing counts until the suite is green. Every hire, handoff and proof is on the console, and the same control plane is exposed over MCP to Claude, Codex or Cursor.
Highlights
- Tests as the bossA task is finished when the test runner says so, then the whole suite runs once more and a proof pack is written.
- Per-task auctionLocal, free and subscription models compete on fit, verified success and cost; a zero budget removes paid routes before the auction starts.
- Cognitive handoffA failed worker leaves a compact checkpoint, 76 to 91 percent smaller than its context, and the replacement resumes instead of restarting.
- Receipts on screenFive modules of a billing SaaS by five workers in 41.6 seconds, whole suite green, paid inference zero, all of it in the execution log.