Your model always sounds sure. Mephistopheles pulls apart every claim it makes, grounds each against independent sources and tools, and tells you what's true, what's wrong, and what it can't verify — with citations.
Every factual statement is pulled out as a standalone, checkable claim — the opinions and fluff left aside.
Each claim is checked against several independent sources, exact tools, and your own trusted documents — not the model's memory.
Supported, contradicted, or unverifiable — each with citations and a confidence, plus one honest risk score for the whole answer.
Point your OpenAI base_url
at Mephistopheles and every answer comes back verified — or call
/verify with any model's output.
Nothing else in your code changes.
# same code, one new line client = OpenAI( base_url = "https://mephistopheles.ai/v1", api_key = "mk-your-key", ) # → replies now carry a per-claim # verdict, citations & risk score
Free to start. One line to add. Every claim accounted for.
Get an API key