Verify LLM API is a tool designed to verify the actual large language model (LLM) powering an AI agent, such as Codex or Claude Code CLI. Its primary function is to run a fingerprint check by sending a series of requests to the agent and analyzing the statistical distribution of the responses. This process does not require extracting API keys or exposing credentials, as it operates through the host's existing login session. The tool is intended for users who need to confirm the identity and consistency of the model behind their agent, rather than relying on self-reported model names or capabilities. The tool operates through a three step workflow initiated by a single prompt. First, it sends a run command to the agent, which downloads a bundled scanner. The scanner then makes a configurable number of fresh requests, defaulting to 20 (4 probe cells with 5 independent requests each), through new, non persistent sessions. Finally, it normalizes the answers, computes the Jensen Shannon divergence between the observed answer distributions and public reference distributions, and prints a report indicating a match, mismatch, or inconclusive result. The scanner automatically selects between Codex or Claude Code and checks only the host's login status without accessing authentication files. Key features include independent sampling across fresh sessions, distribution distance comparison using base 2 Jensen Shannon divergence, and named public references from OpenRouter collected Zenodo datasets. The tool offers three profiles: a quick check with 20 requests, a standard check with 40 requests, and a paper mode with 120 requests. It requires Node.js 22 and a signed in Codex or Claude Code CLI. The tool explicitly states that behavioral distributions are not cryptographic proof or vendor attestation, and that missing references or unstable routes will return an inconclusive result. A separate endpoint scanner for OpenAI compatible and Anthropic endpoints is bundled but only runs when the user explicitly supplies the base URL, model, protocol, and credential variable name.