yohannesjoya/toolreceipt is a Python library that finds claims by LLM agents of having completed work they never actually did. It extracts action claims from the final answer, cross-checks them against the tool calls recorded during execution, and returns completion claims for tools that were never called or that failed as receipts. The repository description presents a record ledger for LangGraph, claim extraction, and receipts as its core. FACTS lists no public release version.
The core verdict is a rule-based approach that uses the standard library. So no separate model call or API key is needed to produce a verdict. It offers inspection functions that can be used standalone, not just the LangGraph integration. Because its structure compares tool call records with the final answer, it focuses on checking what the agent said it did against what was actually recorded. The repository's language is Python and its license is MIT.
However, the README states that it does not support mismatches between claims and tool arguments, quantities, non-English answers, or streaming inspection. So the scope of what this tool verifies is limited to the basic correspondence between recorded calls and the action claims in the answer. It is hard to regard it as a means of verifying claims beyond that scope, and under unsupported conditions a separate check is needed before trusting the results.