konoeph/agentclaimguard

AgentClaimGuard is a Python library that acts as a checkpoint between an AI agent and the user. When an AI language model produces an answer, the library checks whether the key claims in that answer are backed by actual evidence, tool outputs, or other sources you have provided. If a claim lacks the required support, the library blocks it and returns a status indicating why, so the answer can be sent for repair or human review instead of reaching the user unchecked. The library does not judge whether a claim is factually true in an absolute sense. It checks whether a claim meets the rules of a policy you define. For example, if your policy says a numeric conclusion like "revenue increased by 15%" requires a calculator tool result as evidence, and no such result was provided, the library returns a blocked status rather than passing the answer through. The design is meant to close the gap that exists in most AI pipelines, where retrieval gives context but does not guarantee the model used it, and tool calls happen but do not guarantee the model drew the right conclusion. The library is framework-agnostic and can plug into several common setups. It ships with adapters for LangGraph and LangChain, so it can sit as a node inside an existing agent workflow and route outputs to different next steps depending on the verification result. It also includes a FastAPI server that exposes an HTTP endpoint, which lets it be called from workflow platforms like Dify as a plain web request. Installing it is straightforward with pip. Additional optional packages bring in the server, LangGraph adapter, or LangChain adapter depending on which integration you need. The repository includes example scripts for each major use case, including claim extraction helpers that turn plain text answers into structured objects for the verification step, and a mapping pattern for connecting retrieved chunks from a RAG system into the evidence format the library expects. The project is available on PyPI and released under the Apache 2.0 license.