Do I need an API key?

No. Anonymous calls get 3 audits per day per client with up to 10 samples. For more, buy the full audit and send the pass you receive as Authorization: Bearer <pass> .

Why is the anonymous limit so low?

Every audit calls a paid model on our side, and the anonymous tier has no sign-up to stop abuse. The limits keep it free for everyone. Ten samples are enough to see whether a full audit is worth it.

Can my agent pay per call?

Not yet. Pay-per-call over HTTP 402 is planned (see below). Today, the $49 pass is the paid route.

API & MCP

Run the audit from code or an agent

The same audit as the form on the home page, as JSON. Send real inputs with the answers your current model gave; get back agreement, the routing threshold and the cost projection.

Endpoints

POST https://api.llmcutover.com/api/v1/audit: run an audit. Body and response are JSON.
POST https://api.llmcutover.com/mcp: remote MCP server (Streamable HTTP, stateless). Tools: run_cutover_audit, list_models.
GET https://api.llmcutover.com/api/models: model ids, list prices and reasoning-effort levels you can pass as current.model.
OpenAPI 3.1 spec: /openapi.json. Short summary for LLMs: /llms.txt, full text: /llms-full.txt.

Example

curl -s https://api.llmcutover.com/api/v1/audit \
  -H 'content-type: application/json' \
  -d '{"mode":"yesno","question":"Is this support ticket about billing?",
       "current":{"model":"gpt-6.1-sol","effort":"medium"},"usage":{"monthlyCalls":200000},
       "samples":[{"text":"I was charged twice this month","label":"yes"},
                  {"text":"App crashes on login","label":"no"},
                  {"text":"Refund for my last invoice?","label":"yes"},
                  {"text":"How do I export my data","label":"no"},
                  {"text":"Card declined at checkout","label":"yes"}]}'

For extraction, set "mode":"extract", describe up to 12 fields in fields, and give each sample an expected object instead of a label. Samples can also be a CSV or JSONL string.

MCP

Add the server to any client that supports remote MCP over HTTP:

{"mcpServers": {"llmcutover": {"type": "http", "url": "https://api.llmcutover.com/mcp"}}}

With a pass, add the header Authorization: Bearer <pass> in your client’s server settings.

Limits

Caller	Samples per run	Runs	Other
Anonymous API / MCP	5–10	3 per day per client	Shared daily model budget; 64 KB body
Website form	5–25	5 per day	Human check (Turnstile)
Full-audit pass ($49)	up to 500 yes/no or 200 extraction	5 within 30 days	Disagreement list included; 1 MB body

Responses: 429 with Retry-After when you hit your limit, 503 when the shared budget for the day is used up or the service is paused, 413 for an oversized body, 422 with a plain message for bad input. Burst traffic over about 30 requests in 10 seconds from one IP is blocked at the edge for 10 seconds. Calls from cloud and hosting networks get a smaller free allowance.

Data

Samples go to the candidate model’s API for that run and aren’t stored by us. Don’t send personal, health or financial data.

Planned: pay per call (HTTP 402)

Not live. The plan is to answer an over-limit call with 402 Payment Required and a price (for example $0.50 per 100 samples), payable with an agent payment protocol such as x402 or the Machine Payments Protocol, so an agent can buy one more audit without a human checkout. If you need this now, write to contact@llmcutover.com.