Let your AI agent rent a GPU: llms.txt, --json and --budget

Posted by the Lium team. Lium is the GPU marketplace used in this post. Every command below is taken from the live CLI help (lium.io 0.8.0) and lium.io/llms.txt, read on 24 September 2026.

Video (26 s, play the MP4): the opening chat illustrates the ask. The terminal clip is a real recording of lium.io/llms.txt and lium up --help, sped up 1.5x.

Most GPU clouds assume a person with a browser: a sign-up form, a dashboard, a console to click through. An AI agent working in a terminal needs something else: prices it can read without an account, a way to sign up without a form, a rent command that returns machine-readable output, and a hard stop so a forgotten pod can't run up a bill. This post walks through those five steps on Lium, in the order an agent takes them.

1. Read llms.txt

lium.io/llms.txt is a short plain-text file written for models. It says when Lium fits a task (a job that needs an NVIDIA GPU for minutes to days, started from a script) and when it doesn't (managed inference endpoints, serverless functions, CPU-only hosting). It then lists the price feeds, the install commands, the headless sign-up call and the reference docs. One fetch gives an agent everything below.

2. Check prices with no account

curl -sL https://lium.io/pricing.json | jq -r '.models[]
  | select(.available_gpus > 0)
  | "\(.name)  from $\(.min_price_usd_per_gpu_hour)/GPU-hour  \(.available_gpus) free"'

At 02:44 UTC on 24 September 2026 that printed, among others: H100 80GB HBM3 from $1.30 per GPU-hour, H200 from $3.00, B200 from $5.60, B300 from $8.25, RTX 5090 from $0.58 and RTX 4090 from $0.45. Providers set their own prices, so run it again for the current numbers, or see lium.io/pricing. For one row per rentable node, read the public nodes feed.

3. Sign up without a form

curl -f -X POST https://lium.io/api/auth/signup

The response carries a fingerprint (a 32-character dashboard login, shown once, with no recovery) and an api_key when minting succeeds. The key can be null, and sign-up credit depends on the platform setting and on prior use of the IP. The CLI has the same step as lium signup, which stores the key it mints. Renting needs a funded balance that covers at least 15 minutes of the pod's hourly price (billing docs).

4. Rent with JSON output and a budget

lium up --gpu H100 -y --json --budget 12.50 --ttl 3h --timeout 600

What each flag does, from lium up --help:

• --gpu H100 picks the lowest-priced H100 node that matches.
• -y --json skips the prompt, waits for the pod, and prints it as JSON, with no SSH session to hang the agent's shell.
• --budget 12.50 ends the pod once $12.50 has been spent. At $1.30 per GPU-hour that is about 9.6 hours of one H100.
• --ttl 3h ends it three hours after the rent, whichever comes first.
• --timeout 600 gives the whole rent 10 minutes, then exits 1 and names the pod, so the agent can clean up rather than wait forever.

Billing is per second at the node's hourly price, from deploy until the pod is removed, with no minimum. Ten minutes of an H100 at $1.30 per GPU-hour costs $0.22.

5. Run the job, then stop billing

lium exec <pod> "nvidia-smi -L"
lium spend
lium rm <pod> -y

lium exec runs a command in the pod, lium spend shows the hourly burn and the estimated spend per active pod, and lium rm removes the pod, which stops the meter.

Give your agent the skill

The Lium skill documents this sequence and its failure modes for any agent that supports the Agent Skills standard (Claude Code, Cursor, Codex and others):

npx skills add Datura-ai/lium-skill --skill lium
# without node:
curl -fsSL https://lium.io/agents/install.sh | bash

Source: github.com/Datura-ai/lium-skill. The agent guide on the docs site is docs.lium.io/developers/agents.