An LLM API your agent sets up itself.

Tell your agent Set me up on ScaledThought: scaledthought.com/go/org-agents. The link serves /llms.txt, a setup protocol written for agents: every command is non-interactive, prints one JSON object and exits 0, or prints {"error":{"code","message","hint"}} with a code the protocol explains. Agents that can run shell commands, such as Claude Code and Codex, can follow it.

The short version, as the protocol gives it:

• init writes ST_API_KEY and ST_BASE_URL to ./.env (mode 0600) and adds .env to .gitignore. It never prints the key.
• Re-running it is safe: with a key already in .env it creates nothing and returns the existing account.
• Creating an account accepts nothing on the human's behalf. They accept the Terms themselves, on the payment page.
• If the human already has an account, the agent runs npx -y scaledthought@0.3 login --region us instead: the human types a short code on their dashboard, and login --wait writes a new key.
• Only the US region is open. Canada is coming soon.

npx -y scaledthought@0.3 models --live --region us   # what's live; no account needed
npx -y scaledthought@0.3 init --region us   # the account; its key goes to ./.env
npx -y scaledthought@0.3 billing topup --amount-usd 20   # a payment link for the human; charges nothing
npx -y scaledthought@0.3 billing wait   # returns once the human has paid
npx -y scaledthought@0.3 chat -m deepseek-v4-1-flash "Reply with OK"   # verify

Every key can carry scopes, a hard daily budget in USD, rate limits and an expiry of up to 365 days. A sub-agent key (--child) stays within its parent's scopes, budget, rate limits and expiry, and revoking a key revokes every key below it. Keys are stored only as hashes and shown once.

keys put mints a sub-agent key and puts it straight into a platform's secret store (GitHub, Vercel, Fly and others) through that platform's own CLI, with the key on stdin: never printed, never in shell history. The key records where it lives, so a human knows what revoking it breaks.

npx -y scaledthought@0.3 keys create --name researcher --child --budget-usd 2   # a sub-agent key, printed once
npx -y scaledthought@0.3 keys create --name nightly --expires-in 30d --scopes inference,read
npx -y scaledthought@0.3 keys put github --repo acme/app --budget-usd 2   # stored as a GitHub Actions secret
npx -y scaledthought@0.3 keys rotate <id> --grace 24h   # the old key keeps working for the grace period
npx -y scaledthought@0.3 keys revoke-self   # a key that may have leaked revokes itself

Usage is prepaid and there is no subscription. A new account has no credits, so every completion returns 402 insufficient_credits until its human pays. The agent runs npx -y scaledthought@0.3 billing topup --amount-usd 20, which returns an approval_url and opens the funding page in the human's browser. Creating the link charges nothing: only the human can pay, by card on Stripe's checkout, and billing wait returns as soon as they have.

Top-ups are $10 to $2,000 per top-up, at most $2,000 per 24 hours and 2 open approval links at a time. The human can also turn on auto-reload on the funding page. An agent can turn auto-reload off or lower it, but turning it on or raising it always goes back to the human (403 approval_required). Prepaid credits and budgets

API responses carry x-st-notice: low_balance once the credit balance is under $5. The agent tells its human what it means and what to approve, instead of running until requests fail. GET /v1/account, the MCP tool account_status and the CLI's doctor show the same notices.

Instead of polling, an agent can register a webhook for balance.low, key.expiring and budget.exhausted. Deliveries are signed (x-st-signature, HMAC-SHA256) and carry ids, amounts and timestamps only, never prompts.

npx -y scaledthought@0.3 webhooks add https://hooks.example.com/st --events balance.low,budget.exhausted

The same account works three ways. The CLI (npx -y scaledthought@0.3) prints JSON whenever its output isn't a terminal. The HTTP API creates an account with POST /v1/accounts (no auth; the key is returned once) and manages keys, budgets, usage and billing. The MCP server at https://api.us.scaledthought.com/mcp gives an MCP client account tools such as account_status, keys_manage, budget_manage, usage_query and billing_topup; it can't create accounts, so sign up first.

Inference is the OpenAI-compatible API: chat completions, the Responses API and batches, with any OpenAI SDK. Point it at ST_BASE_URL and ST_API_KEY from ./.env.

import os
from openai import OpenAI

client = OpenAI(base_url=os.environ["ST_BASE_URL"], api_key=os.environ["ST_API_KEY"])
resp = client.chat.completions.create(
    model="deepseek-v4-1-flash",  # or a model id from `npx -y scaledthought@0.3 models --live`
    messages=[{"role": "user", "content": "Hello"}],
)

• Paying. Only the human pays, on Stripe's page, and accepts the Terms there.
• Turning auto-reload on or raising it.
• The dashboard (billing, keys, teammates, connecting agents), which the agent opens for them with npx -y scaledthought@0.3 dashboard. A person can also sign up there themselves, by email link: app.us.scaledthought.com/sign-up (US accounts only for now).