Launch AI experiments on cloud GPUs.
Catapult snapshots your project, runs it on a cloud GPU, streams logs and metrics back to your terminal, and brings your checkpoints home. Set a price and a time cap, and pay only for the seconds you use.
aws · lambda · skypilot · one launch
$ origami run train.py --gpu a100 --max-price 1.50 --max-runtime 4h
✓ snapshot 142 files (3.2 MB)
✓ provisioning a100 on lambda · $1.10/hr
✓ streaming logs
epoch 12 · loss 1.84 · val/acc 0.71
✓ downloaded checkpoints/ → ./runs/2f9c1
The launch sequence
Snapshot, launch, stream, download
Snapshot
Catapult snapshots your code and data so the run is reproducible — no manual rsync, no 'works on my machine'.
Launch
It provisions a GPU across providers and starts your command, picking the cheapest instance that fits your caps.
Stream
Logs and metrics stream straight to your terminal while the run executes, so you watch it like a local process.
Download
Checkpoints and outputs sync back to your machine automatically when the run finishes — or when you stop it.
Features
Everything a run needs, nothing it doesn't
Spend caps
Set --max-price and --max-runtime. Catapult stops the run before it overspends, so a forgotten job never becomes a surprise bill.
Multi-cloud
AWS, Lambda, and SkyPilot behind one command. It launches on whichever has capacity at the lowest price.
Safe snapshots
Your project is snapshotted before each run, so what executed on the GPU is exactly what you can reproduce later.
Live logs
stdout and stderr stream to your terminal in real time. Tail a remote run as if it were running locally.
Artifacts
Checkpoints, logs, and outputs are pulled back to your machine automatically — nothing stranded on a dead instance.
Workspaces
Persistent environments carry your dependencies and data across runs, so you don't rebuild from scratch each time.
Experiments & runs
Runs are grouped into experiments you can list, compare, and re-launch instead of digging through shell history.
MCP
Drive Catapult from your agent or editor over MCP — launch and inspect runs without leaving your tooling.
Pricing
Prepaid credits. Pay for the seconds you use.
Top up with credits and spend them by the second of GPU time. No seats, no monthly minimum, no idle charges. Your --max-price and --max-runtime caps mean a run can only ever cost what you allowed.
Prepaid credits
Pay as you go
- Billed per second of GPU time
- Hard spend caps on every run
- No seats or monthly minimums
- Credits never expire
Load it up. Let it fly.
Install the CLI, point it at your script, and set your caps. Catapult handles the rest.