In recent years, Large Language Models (LLMs) have demonstrated potential in vulnerability detection works, yet their practical application has been hindered by limited contextual processing and complex task analysis capabilities. In this work, we propose VulPelican, a novel LLM and interactive static analysis tool based vulnerability detection framework. VulPelican implements multi-round interactions through the ReAct framework: First, the LLM analyzes code and generates its query to the tool, then Joern retrieves the code repository to supplement contextual information such as cross-file data flows and function dependencies, with results feed back to the LLM for further analysis. Evaluations on the InterPVD dataset demonstrate that VulPelican achieves significantly superior overall accuracy (0.688) and F1-score (0.657) compared to traditional single-round zero-shot LLM vulnerability detection methods (0.472 and 0.552), with glm-4-flash and deepseek-v3 showing optimal performance when being base models. By fusing LLMs’ semantic generalization capabilities with the deterministic paths of static analysis, VulPelican provides an innovative solution for detecting complex vulnerabilities, validating the technical feasibility of interactive frameworks in intelligent code auditing.