app_check
This tool on every surface: CLI · MCP · Python · JavaScript · Go · Java · .NET
Check a claim about a recorded session: a model with a fresh context — none of yours — reads the recording (each call, its result or error, and the screen and screenshot after it) through read-only tools, and answers passed, failed or inconclusive with the observations it rests on. Inconclusive means the recording cannot settle it; nothing missing is taken for success. A model's judgment is not proof: keep deterministic assertions (app_wait_for) for what can be asserted directly, and use this for claims about what a person would have seen. Needs no device. The default judge, Claude through the Anthropic API, needs an Anthropic credential and each check is a paid call; it is built but not yet measured against claims with known answers. Judges through "openai:
Read-only: readOnlyHint is true.
Arguments
| Name | Type | Required | What it is |
|---|---|---|---|
claim | string | yes | What should be true, precisely: "the terms box was ticked after the tap". |
model | string | The model that judges. Defaults to claude-opus-5-5, not yet measured. "openai: | |
record | string | yes | A recording: the .zip app_trace with action stop saved. |
Examples
Each is the arguments of a tools/call request; the first is shown whole.
Ask a model on this machine whether a claim holds in a recording
{
"jsonrpc": "2.0",
"id": 1,
"method": "tools/call",
"params": {
"name": "app_check",
"arguments": {
"claim": "After card 3 was tapped, it was shown face up as the Blue Morpho.",
"record": "card.zip",
"model": "openai:qwen2.5:14b"
}
}
}