How it works
Five moving parts. You only ever touch the first one.
The Relay Pipeline
Devs drop jobs (code builds, model runs) with deadlines & privacy tags.
Evaluates rate limits, privacy restrictions, cost, and machine idle stats.
Backfills paid subscription quota before it resets.
Harvests idle horsepower at zero marginal cost.
Add units of work to one queue — a coding task, a batch job, a model run. Tag each with priority, a deadline, and a privacy level.
The router scores every eligible endpoint by privacy fit, cost, latency, and free capacity, then assigns the job to the first worker that qualifies.
Relay watches remaining subscription quota and the reset clock, then backfills idle paid capacity with deferrable work before it expires.
When a local machine goes idle, its runner pulls free, eligible jobs from the queue — and yields the instant you start using the machine again.
One board shows throughput, spend, and anything stuck. Failed jobs retry or fail over to another endpoint automatically.
Routing rules
| Job looks like | Relay sends it to | Because |
|---|---|---|
| Private source code | Local runner | Privacy tag forbids leaving the machine |
| Huge context, not urgent | Cloud, backfilled | Uses paid quota that would expire anyway |
| Small + overnight | Idle local machine | Free compute you already own |
| Urgent + heavy | Fastest available cloud | Deadline beats cost |
Let the queue keep every brain busy while you build the next thing.