What is provena?

Provena is a Python library that keeps track of where an AI agent's information came from, and whether it can be trusted. When an AI agent pulls in data from several different sources, such as a search tool, a database, or another agent, it becomes hard to answer basic questions afterward: which source provided which piece of information, was any of it changed after the fact, and was it still up to date when it was used. Provena answers those questions by wrapping the functions that produce that context and logging every call. Each logged entry gets a cryptographic hash and is chained to the ones before it, so if anyone tries to tamper with the log afterward, the chain breaks and the tampering becomes detectable. Provena can also check whether context carries proper source information and flag it as missing or incomplete, and it can check whether the context is still fresh or has gone stale based on timestamps. On top of just observing, it can enforce rules: it can block, warn about, or simply log context that fails these checks before that context reaches the AI model. The core library has no dependencies beyond Python itself and adds very little processing overhead. Optional add-ons bring in a command-line tool, PostgreSQL storage, a server for the Model Context Protocol, PDF compliance reports, and ready-made connectors for popular agent frameworks such as LangChain, LlamaIndex, CrewAI, AutoGen, the OpenAI Agents SDK, and Google's ADK. It can also track handoffs between multiple agents and summarize gaps across all of them, such as missing source information or broken audit chains. This is aimed at developers building AI agent systems, especially ones subject to compliance requirements like the EU AI Act, who need to prove what information their agents used and that it was not altered. It installs with a single pip command and is released under the Apache 2.0 license.