What are verifiable AI tool calls?

Verifiable AI tool calls are agent tool invocations with a public execution receipt that records the call, observed result, and integrity fingerprint.

What is the difference between verifying an agent and verifying a tool call?

Verifying an agent concerns the identity or behavior of the agent, while verifying a tool call checks the specific recorded invocation, arguments, result, and receipt fingerprint.

How do I verify an AI agent tool call?

Open the receipt URL, check the named tool and arguments, then recompute the SHA-256 fingerprint against the observed result.

DISAMBIGUATE / DEFINE / VERIFY

Verifiable AI Tool Calls

Verifiable AI tool call definition

Verifiable AI tool calls are agent tool invocations that mint a public execution receipt: what ran, with what arguments, what was observed, and a fingerprint binding the record to that outcome.

Agent tool call verification has two kinds: before execution, checking the call against a policy or manifest; and after execution, checking the receipt that proves the call happened as recorded.

To verify an AI tool call after it runs, open the receipt URL, check the named tool and arguments, and recompute the SHA-256 fingerprint against the observed result.

AI phone calls vs AI tool calls

AI calls can mean voice or phone calls, which voice platforms verify at the call-record layer. This page covers agent tool invocations: requests from an AI agent to a connected capability, with a public receipt for the recorded execution.

Live examples

live_price tool callOpen the public execution receipt and inspect its recorded result. capability_search tool callCompare a second public receipt from the same reference implementation.

Two kinds of verification

Before execution

Check the proposed tool, arguments, permissions, and policy or manifest before allowing the call.

After execution

Open the receipt, inspect the recorded call and result, and recompute its SHA-256 fingerprint.

How to verify an agent tool call

Open the receipt URL and confirm the record identity.
Check the named tool and the arguments that were recorded.
Read the observed result and execution status.
Recompute the SHA-256 fingerprint against the observed result.

Read the execution receipt definition, the MCP execution receipt definition, and the MCP documentation.

Execution receipts MCP receipts MCP docs