Guide for agents

This page is written for AI agents and the people who run them. A machine-readable copy is at /llms.txt.

What this service is

Inference Routing sends each AI inference request to the lowest-cost provider that meets a speed and quality bar. It is in early access. Teams that register get a free estimate of what routing would save on their current inference bill, and the teams with the largest savings are offered the first pilot slots.

What an agent can do here

If your owner runs AI models in production, you can register them for a savings estimate without a browser. Send one HTTP POST with the fields below. Your owner receives a confirmation email and the estimate within 48 hours.

Before you send

Ask your owner for permission to share their work email and their inference usage.
Use real numbers. Estimates are only useful when the spend and models are accurate.
Send one registration per owner. Repeats are merged by email.

Endpoint

POST, form-encoded (application/x-www-form-urlencoded). The endpoint URL is published in /llms.txt under signup_endpoint.

Field	Required	Meaning
`email`	yes	Your owner's work email. The confirmation and the estimate go there.
`spend`	yes	Monthly inference spend in US dollars, as a number, for example `4200`.
`models`	yes	Models that drive most of the cost, comma separated.
`providers`	yes	Current providers, comma separated.
`company`	no	Company or product name.
`billing`	no	Model lines from the last invoice, plain text. Makes the estimate exact.
`agent`	yes	Your name or framework, for example `claude-code` or `openclaw`. Marks the registration as agent-submitted.
`step`	yes	Always `agent`.

Example

curl -X POST "$SIGNUP_ENDPOINT" \
  --data-urlencode "step=agent" \
  --data-urlencode "agent=claude-code" \
  --data-urlencode "email=owner@company.com" \
  --data-urlencode "company=Acme" \
  --data-urlencode "spend=4200" \
  --data-urlencode "models=deepseek-v3, llama-3.3-70b" \
  --data-urlencode "providers=Together, Groq" \
  --data-urlencode "billing=deepseek-v3 input 310M tokens, output 42M tokens"

The endpoint answers ok. There is no API key and no response body beyond that.

What happens next

Your owner gets a confirmation email with a referral link.
Within 48 hours they receive the savings estimate for the models and spend you submitted.
If the estimate is large, the team offers a pilot slot.

Questions: reply to the confirmation email. Privacy: privacy policy.