How do I give an AI agent a spending limit?

Create a key with a budget, orl api-keys create --name agent --budget-usd 25 --budget-period daily, and run the agent with it as OPENRELAY_API_KEY. Budgets run per UTC day, per UTC month or for the key's lifetime. A key with a budget cannot create keys, so the agent cannot mint itself one without a limit.

Can each team get its own budget?

Yes. Fund the organization once and give each team or agent its own key with a budget, monthly or for the key's lifetime: every key draws on the same prepaid balance and stops at its own limit. A key budget counts inference and batches, not GPU rentals, so GPU spend stays bounded by the organization's balance.

What happens when the budget runs out?

The key's next inference request or batch is refused with 402 key_budget_exceeded before it reaches a model. Daily and monthly budgets reset at 00:00 UTC. The check uses each request's estimated cost and reserves nothing, so requests already running, or one large batch, can pass the limit.

Does the key budget cover GPU rentals?

No. A key budget counts inference and batch spend. VMs and pods bill to the organization's balance, so for a hard limit on everything an agent spends, give it its own organization with a prepaid balance and auto-recharge off: running workloads stop when the balance reaches the $0.50 floor.

Can an agent look without spending?

Yes. orl mcp serve --read-only gives the agent only the read tools, so it can list models, VMs and usage but cannot create anything. For a limit that holds outside MCP too, give the agent a key with fewer scopes.

How does the agent pay?

From your organization's prepaid balance, which you fund with a card. The agent never holds a payment method: it holds an API key, and the key's budget bounds what it can draw on inference.

Inference and GPUs for AI Agents

Let your agents call models and rent GPUs on their own, through orl or the API. Each one gets an API key with a spend budget, so its inference stops at the limit you set.

Add a card and make a deposit to start.

What agents can run.

Models, batches and GPU VMs, on the same key.

Inference pricing

Inference · per token

Call a model

Every response says what it cost.

From $0.045 / 1M in

See models

Batch · 50% off

Run a batch

JSONL in, results out, at half the on-demand rate.

Evals
Backfills

Batch for agents

GPU VMs · per minute

Rent a GPU

A retry with the same idempotency key returns the first VM.

H100 $2.60/hr

See GPU pricing

Spend stops where you set it.

The agent holds a key, not a card.

Key budgets

Per call · $1 by default

A cap on every call

A chat or batch call whose cost bound is over the cap never runs.

Chat
Batch

Per key · day, month or lifetime

A budget on the key

One prepaid balance, a key and a budget per team or agent. Past its limit, a key gets a 402.

Inference
Batch

Per organization · prepaid

A floor under everything

With auto-recharge off, workloads stop at the $0.50 floor. GPUs included.

Inference
Batch
GPUs

What a budget buys.

Requests a day at the catalog's rates when this page loaded.

Inference pricing

Daily budget$25 / day

Model

Workload16K in · 1.5K out

Requests a day

Each bills $0.0377 on GLM 5.2, so $25 a day moves 11.6M tokens.

Prompt 77%Output 23%

CLI or API.

One key either way, and its budget holds on both.

orl, from a shell or MCP.

JSON out, stable exit codes, no prompts. orl mcp serve gives Claude Code and Cursor the same commands as tools.

orl for agents

orl cli

$ brew install openrelayinc/tap/orl
$ orl auth login
$ claude mcp add openrelay -- orl mcp serve
✓ Added stdio MCP server openrelay

The API, from any SDK.

Inference speaks the OpenAI and Anthropic formats at inference.openrelay.inc. VMs, pods and keys live at api.openrelay.inc.

API reference

api

$ curl https://inference.openrelay.inc/v1/chat/completions \
    -H "Authorization: Bearer $OPENRELAY_API_KEY" \
    -d '{"model": "openrelay/gpt-oss-120b", ...}'
{ …, "openrelay": { "cost_nano_usd": 744600 } }

Agents, answered.

Talk to the team

Yes. Over MCP it calls the vms_create tool; from a shell it runs orl deploy or orl vms create. The VM bills to your organization's balance. Pass an idempotency key and a retried create returns the first VM instead of renting a second.

Two ways. As tools: claude mcp add openrelay -e OPENRELAY_API_KEY=<the agent's key> -- orl mcp serve lets Claude call models, run batches and deploy GPU VMs, with each chat or batch call capped at $1 unless you raise --max-cost. As its model: set ANTHROPIC_BASE_URL=https://inference.openrelay.inc and ANTHROPIC_AUTH_TOKEN to the key, and pick a model such as openrelay/glm-5.2. For Cursor, orl init --agent --mcp cursor writes the MCP server into .cursor/mcp.json.

Give your agent a key.

Create its key with a budget, then point the agent at orl.

Inference and GPUsfor AI agents.