A Secure Pseudo-Hosted Framework for LLM Conversations in Virtual Environment

Abstract

A privacy-preserving approach for interacting with Large Language Models (LLMs) using a pseudo-hosting methodology within a terminal-based environment. In today’s world, most LLMs are accessed online through browsers or hosted cloud services, where user data such as prompts, responses, and metadata are often stored or tracked. This raises significant privacy concerns, including risks of data leaks, session hijacking, and unauthorized profiling. Meanwhile, local hosting offers privacy but demands powerful hardware, large memory allocation, and complex configurations. To bridge the gap between these two extremes, this paper proposes a pseudo-hosted LLM integration, where the model runs directly inside a terminal within a virtual environment (venv), ensuring isolation, data protection, and low system resource usage. The implementation uses Google’s Gemini API key for establishing secure communication with the LLM, while dotenv is utilized to protect sensitive credentials by storing them in environment variables. The requests module facilitates query handling between the user and the model, ensuring smooth and efficient data exchange. For user convenience, the chatbot supports both text and voice-based interactions using speech_recognition for converting speech to text and pyttsx3 for converting text responses back into speech, providing a complete conversational experience through the command-line interface (CLI). Once the session ends, all chat data is automatically erased, ensuring 0% data retention and no history tracking, unlike typical online AI chat platforms. The pseudo-hosting methodology provides several advantages: it ensures enhanced privacy and data isolation, requires minimal hardware resources (no GPU, NPU, or large RAM), and functions entirely offline except for API requests. Additionally, by operating inside a virtual environment, it prevents any external scripts, cookies, or cached elements from accessing or recording user activity. Although it lacks a graphical interface and depends on API quotas, this approach successfully combines the privacy of local hosting with the ease and accessibility of online AI, forming a balanced, efficient, and secure LLM execution framework. The proposed system demonstrates that advanced LLMs can be integrated in a safe, lightweight, and privacy-conscious manner—marking a step forward in responsible AI usage and user data protection.

Introduction

Large Language Models (LLMs) have transformed digital communication, powering chatbots that assist in education, research, and professional workflows. However, most LLMs operate through cloud-based platforms that store user prompts and responses on external servers, raising serious privacy and data security concerns. Issues such as data retention, unauthorized access, and lack of transparency in model training have intensified debates around ethical AI use.

To address these risks, this study proposes a terminal-based pseudo-hosted LLM interface—a lightweight, privacy-focused alternative to traditional web-based systems. Unlike cloud chatbots that log interactions, the terminal environment (e.g., within Visual Studio Code) operates as an isolated, ephemeral shell that stores no cookies, history, or cached data. Once the session ends, all conversation traces are permanently erased, ensuring non-persistence and user autonomy.

This pseudo-hosting model combines the privacy benefits of local deployment with the convenience of cloud APIs, enabling secure LLM interaction without heavy hardware requirements. Implemented within a Python virtual environment, it uses lightweight modules (e.g., google.generativeai, speech_recognition, pyttsx3) for flexible, text- and voice-based communication.

The proposed system follows a stateless architecture: user queries are sent to an API, processed, displayed, and then discarded. No logs are kept locally or remotely, achieving complete data ephemerality. Applications include privacy-sensitive academic research and secure professional workflows, where confidentiality is crucial.

Ultimately, the paper presents a hybrid, privacy-preserving framework that bridges the gap between usability and data protection. By prioritizing ephemeral communication, minimal data persistence, and ethical AI deployment, the terminal-based LLM model offers a practical step toward secure and responsible human–AI interaction.

Conclusion

The implementation of the pseudo-hosted terminal-based LLM interface demonstrates a significant advancement in balancing usability, privacy, and accessibility for everyday users. The vast majority of people leverage AI to complete tasks efficiently and effectively, with seven out of ten individuals using AI in their daily routines, yet only a small fraction of users actively considers privacy concerns. Traditional AI platforms often lack transparency regarding user data processing, highlighting a critical gap between convenience and data security. The proposed approach addresses this gap by running the LLM within a virtual or isolated environment, akin to VS Code’s Shell and virtual environment variables, allowing the model to function fully while ensuring that user data remains local, ephemeral, and secure. This design allows users to interact with AI without exposing their conversations to potential session hijacks or unauthorized storage, though it comes with minor trade-offs, such as the absence of a graphical user interface and the need for some manual configuration. Despite these compromises, users still gain full functionality and output quality comparable to cloud-hosted AI, but with complete control over privacy. Furthermore, the pseudo-hosting methodology opens avenues for future enhancements, including the integration of diffusion models or custom server-based deployment, which would allow even more sophisticated AI interactions while maintaining privacy and offering UI customization. While more computationally intensive models like diffusion models may require external GPUs or additional VRAM for optimal performance, the current system demonstrates that a secure, low-spec, terminal-based LLM can serve as a practical, privacy-conscious alternative for a wide range of users, bridging the gap between convenience, efficiency, and data security.

References

[1] M. Kahng et al., \"LLM Comparator: Interactive Analysis of Side-by-Side Evaluation of Large Language Models,\" in IEEE Transactions on Visualization and Computer Graphics, vol. 31, no. 1, pp. 503-513, Jan. 2025. [2] D. S. Johnson, O. Hakobyan, J. Paletschek and H. Drimalla, \"Explainable AI for Audio and Visual Affective Computing: A Scoping Review,\" in IEEE Transactions on Affective Computing, vol. 16, no. 2, pp. 518-536, April-June 2025. [3] J. Berengueres, \"How to Regulate Large Language Models for Responsible AI,\" in IEEE Transactions on Technology and Society, vol. 5, no. 2, pp. 191-197, June 2024. [4] L. Benaddi, C. Ouaddi, A. Jakimi and B. Ouchao, \"A Systematic Review of Chatbots: Classification, Development, and Their Impact on Tourism,\" pp. 78799-78810, vol. 12, in IEEE Access, 2024. [5] D. Noori Zadeh and M. B. Elamien, \"Generative AI for Analog Integrated Circuit Design: Methodologies and Applications,\" pp. 58043-58059, vol. 13, in IEEE Access, 2025. [6] S. Iftikhar, S. H. Alsamhi and S. Davy, \"Enhancing Sustainability in LLM Training: Leveraging Federated Learning and Parameter-Efficient Fine-Tuning,\" in IEEE Transactions on Sustainable Computing,2024. [7] S. Karnouskos, \"The Relevance of Large Language Models for Project Management,\" pp. 758-768, vol. 5, in IEEE Open Journal of the Industrial Electronics Society, 2024. [8] Q. Lu, L. Zhu, X. Xu, Z. Xing and J. Whittle, \"Toward Responsible AI in the Era of Generative AI: A Reference Architecture for Designing Foundation Model-Based Systems,\" pp. 91-100, vol. 41, no.6, in IEEE Software, 2024. [9] T. Scantamburlo et al., \"Software Systems Compliance with the AI Act: Lessons Learned from an International Challenge,\" pp. 44-51, in IEEE/ACM International Workshop on Responsible AI Engineering (RAIE), Lisbon, Portugal, 2024. [10] A. Koubaa, W. Boulila, L. Ghouti, A. Alzahem and S. Latif, \"Exploring ChatGPT Capabilities and Limitations: A Survey,\" pp. 118698-118721, vol. 11, in IEEE Access, 2023. [11] O. Mubin, F. Alnajjar, Z. Trabelsi, L. Ali, M. M. A. Parambil and Z. Zou, \"Tracking ChatGPT Research: Insights from the Literature and the Web,\" pp. 30518-30532, vol. 12, in IEEE Access, 2024. [12] A. T. Neumann, Y. Yin, S. Sowe, S. Decker and M. Jarke, \"An LLM-Driven Chatbot in Higher Education for Databases and Information Systems,\" pp. 103-116, vol. 68, no. 1, in IEEE Transactions on Education, 2025. [13] P. Haindl and G. Weinberger, \"Students Experiences of Using ChatGPT in an Undergraduate Programming Course,\" pp. 43519-43529, vol. 12, in IEEE Access, 2024. [14] S. Yu, F. Carroll and B. L. Bentley, \"Insights Into Privacy Protection Research in AI,\" pp. 41704-41726, vol. 12, in IEEE Access, 2024. [15] M. Amin Kuhail, M. Bahja, O. Al-Shamaileh, J. Thomas, A. Alkazemi and J. Negreiros, \"Assessing the Impact of Chatbot-Human Personality Congruence on User Behavior: A Chatbot-Based Advising System Case,\" pp. 71761-71782, vol. 12, in IEEE Access, 2024.

Copyright

Copyright © 2025 Krithiga B, Kavidha A, Marikkannan M. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.