Everything the app does, your product or agent can do: measure how AI answers name a company, have the fixes written and checked, and read a holdout lift verdict. $0.015 per measured answer, with no monthly minimum.

Holdout has not launched yet. The API and MCP server are built, but there is no public address and no way to get a key today. Key sign-up is coming soon.

The same four steps a customer takes

Check a website. POST /v1/audits (tool create_audit). It returns at once. Read the audit until it is done.
Attach it to your plan. POST /v1/workspaces (tool create_workspace). The first full measurement starts and the holdout split is frozen.
Connect the site, once. POST /v1/workspaces/{id}/connections (tool connect_site). Checked fixes are then written and published with no further call, while Autopilot is on.
See what changed, and undo one change. GET /v1/workspaces/{id}/autopilot (tool list_activity) and POST /v1/workspaces/{id}/fixes/{fixId}/undo (tool undo_fix).

Buying the first plan is a step for a person, at checkout. A key only exists once an account has a plan, so an agent with a key always acts inside a plan someone agreed to.

Keys and scopes

Add an Authorization: Bearer header to every request. Keys are shown once and stored only as a hash. Give each key the least it needs.

read: Read audits, stats and fixes.
measure: Start audits and measurement runs. Billed per measured answer.
approve: Approve drafted fixes. Connecting a site, changing the mode, pausing, approving and undoing all need this scope, because they can change a website once it is connected.

Endpoints

Every MCP tool has the same name and does the same thing as the endpoint beside it.

Endpoint	MCP tool	Scope
`POST /v1/audits`	`create_audit`	`measure`
`GET /v1/audits/{id}`	`get_audit`	`read`
`POST /v1/workspaces`	`create_workspace`	`measure`
`GET /v1/workspaces`	`list_workspaces`	`read`
`POST /v1/workspaces/{id}/connections`	`connect_site`	`approve`
`POST /v1/workspaces/{id}/implement`	`implement_fixes`	`approve`
`GET /v1/workspaces/{id}/autopilot`	`list_activity`	`read`
`POST /v1/workspaces/{id}/autopilot`	`set_autopilot`	`approve`
`POST /v1/workspaces/{id}/autopilot/pause`	`pause_autopilot`	`approve`
`POST /v1/workspaces/{id}/fixes/{fixId}/undo`	`undo_fix`	`approve`
`GET /v1/workspaces/{id}/fixes`	`list_fixes`	`read`
`POST /v1/workspaces/{id}/fixes/{fixId}/approve`	`approve_fix`	`approve`
`POST /v1/workspaces/{id}/measure`	`measure_workspace`	`measure`
`GET /v1/workspaces/{id}/stats`	`get_stats`	`read`
`GET /v1/workspaces/{id}/ai-visits`	`get_ai_visits`	`read`

Safe retries

Add an Idempotency-Key header to any POST. A repeat with the same key and body returns the first response, and nothing runs or is billed twice. The same key with a different body is refused. Long work (audits, measurements, publishing) returns at once and runs in the background, so a timeout never leaves you guessing. Read the resource again.

Errors

Every error has a stable code, a sentence a person can read and, where it helps, a hint that says what to do next.

{
  "error": {
    "code": "missing_scope",
    "message": "This key does not have the \"approve\" scope.",
    "hint": "Create a key with that scope in Settings."
  }
}

Metering

You pay for measured answers only. One measured answer is one question asked once on one engine. An audit is 16 answers. A workspace measurement is questions times engines times repeats. Reading, connecting and undoing are free, and so is every fix we make to your site.

MCP server

A stateless MCP server over streamable HTTP, with the same key and scopes as the REST API. It has 15 tools. Tools that only read are marked read-only, so a client can let an agent call them without asking.

An agent can run the whole service for a site: check it, attach it, connect it and let the fixes ship. It can also pause Autopilot and undo any change. It cannot message other websites. No key can, by design.

What the numbers mean

Every rate comes with its count and a 95% Wilson interval. There is no rank and no single-run result. Each response names the engines and the access method, because answers from an API can differ from the consumer app. The lift verdict is one of lift_detected, no_lift_detected or not_enough_data, with the number of further runs needed. See the method.