nsankar/aletheia

Aletheia is an AI investigator built to check whether a claim is actually true, rather than just summarizing whatever search results sound the most confident. You might ask it something like whether a company really has the revenue or customer count it claims, and instead of taking the loudest answer at face value, it treats the true answer as hidden and every search result as a noisy clue that only partly reveals it. The core idea is what the project calls an uncertainty loop: form a belief about what is likely true, decide which single piece of evidence would change that belief the most, go look for it, then update the belief based on what was found, including letting contradicting evidence lower its confidence. It keeps repeating this until it is both confident and settled, or it honestly reports that the evidence was inconclusive rather than guessing. The output is a verdict with a plain-English confidence level, the evidence and sources behind it, and whatever remains unresolved. Aletheia runs entirely on your own machine. There is nothing to upload or publish, no required dependencies, and even Python itself is optional since it can fall back to basic arithmetic without it. It is designed to work inside Claude Code or OpenAI Codex, using whichever search tools that agent tool already has available, and support for Claude Cowork is listed as a work in progress. To use it with Claude Code, you clone the repository and copy its skill folder into your local Claude skills directory, then simply ask a plain question in a new chat, since it engages based on intent rather than being called by name. With OpenAI Codex, you point it at the included AGENTS.md file instead. While the shipped examples focus on checking claims about companies, such as vendors or competitors, the README notes the same approach could be pointed at other domains where evidence is noisy and a claim needs checking, like verifying a contractor's credentials or checking whether a viral science headline matches what a study actually found. Doing so mainly means changing what questions it tracks and which sources it trusts. The project is licensed under the MIT License.