A GPU for every agent.
Claude Code, Codex, and Cursor can train, evaluate, and render on remote hardware — inside limits you set once. Your keys never leave your machine, and nothing stays running after the command ends.
curl -fsSL https://downloads.amics.ai/cli/install.sh | shYou approve the budget. The agent does the rest.
Install the skill and your agent knows how to ask for hardware, respect the cap, and clean up after itself. You don't pick a region, watch a dashboard, or remember to shut anything down.
Two things cross the wire
A snapshot of the current revision, and the one command your agent was approved to run. Everything else — the part that would actually hurt to leak — never leaves your disk.
This is not SSH, a remote desktop, or a mounted filesystem. It runs one command and gives you back its exit code.
How the snapshot is builtStays on your machine
- Agent session, prompts, and tool calls
- Branches, uncommitted edits, git history
- SSH keys and cloud logins
- Model API keys
Crosses the wire
- A snapshot of this revision, scanned for secrets
- The literal command, arguments unchanged
- Only the env values you name with
--env
That's the whole payload.
Everything your agent ran is on the record
Every run shows up here with its hardware, region, cost, and exit status. Runs are durable, so a dropped connection doesn't lose the work — and an interrupted run is held for you to inspect, never silently retried.
Compute
GPU clusters you have provisioned.
| Cluster | Status | GPU | Region | Price |
|---|---|---|---|---|
| train-autoresearchsky-l4-1x-2f1a | Active | L4 x1 | United States | $0.36/h |
| eval-sweepsky-a100-1x-77c3 | Queued | A100 x1 | Europe | $1.23/h |
| nightly-finetunesky-h100-1x-90ab | Succeeded | H100 x1 | United States | $1.42/h |
Nothing runs when you aren't using it
No always-on machine, no monthly commitment, no reserved capacity to forget about. You're billed for the time the run is up, and a short job costs a few cents.
L4
24 GB VRAM
from $0.36/hr
Fine-tunes, evals, batch inference
A100
80 GB VRAM
from $1.23/hr
Mid-size training, long context
H100
80 GB VRAM
from $1.42/hr
Serious training runs
Rates and billing granularity follow the underlying capacity, so the CLI quotes you the real price before anything launches and never exceeds the ceiling you set.
Guardrails, not good intentions
These aren't instructions in a prompt your agent might reason its way around. They're enforced by the CLI and the control plane.
- Asking what a GPU costs is read-only. Looking never spends.
- An unattended run won't start without a price ceiling and a deadline.
- A quote above your ceiling never launches.
- Credentials are never forwarded, and the snapshot is scanned for secrets.
- Machines acquired for a run stop when the command ends.
- An interrupted run is held for inspection, never silently re-run.
76% below your ceiling
- GPU
- L4 · 24 GB
- Your ceiling
- $1.50/hr
- Deadline
- 30 minutes
- Warm after exit
- off
Then hand it to your agent
Sign in once, install the skill, and your agent learns the whole workflow — how to ask for hardware, respect the cap, and clean up.
amics login
npx skills add amics-ai/skills --skill amicsStart it on something small
Give your agent a short, restartable job and a ceiling you're comfortable with. Check what it cost when the exit code comes back.