Verdict reviews the design, the copy and the brief the way a sharp colleague would: it finds what a user would trip over, says how sure it is, and leaves the decision with you. Six tools, one endpoint, built so an agent can call them before it hands work to a person.
One per line, as type: string. Types: button, error, empty_state, confirmation, notification.
Runs on the free demo allowance: 2 calls a day, no signup. Your text goes to the review model and is not stored.
Give it a screenshot URL or a description, and a brief if you have one. It comes back with findings ranked by how badly they hurt a real user, each marked as observed or inferred, plus the questions most designs skip: first use, the 500th use, no data, too much data, no permission, no undo.
The states a feature can be in, and the ones nobody drew. Includes the billing, outside-service and two-sources-of-truth cases.
The stages around an objective, plus every arrival and exit the flow forgot.
Buttons, errors, empty states and notifications, string by string, with a fix for each real problem.
The beliefs inside a brief that nobody checked, each rewritten as a question you could answer this week.
One concept under three names across your UI, docs and API.
As an MCP server:
{
"mcpServers": {
"verdict": {
"url": "https://verdict.allthepossibles.com/mcp"
}
}
}
Or over REST, one route per tool:
curl -X POST https://verdict.allthepossibles.com/api/audit_copy \
-H "content-type: application/json" \
-d '{"strings":[{"text":"Submit","context":"button"}]}'
All six tools, 2 calls a day per caller, no key. It is a demo budget and can close early on a busy day. Send your own Anthropic key in an x-anthropic-key header and the limit goes away, because the call runs on your account.
A key, a higher rate limit and a monthly review allowance.
Five times the allowance of Basic and the highest rate limit, for CI jobs and agent loops that review everything they produce.