FinAgent: A Privacy-Preserving Multi-Agent Financial Intelligence System with Context-Aware Governance and GraphRAG Reasoning

Abstract

This paper proposes FinAgent, a new privacy-preserving, local-first multi-agent financial intelligence system that tackles key challenges in automated financial analysis: context management, multi-hop reasoning, and explainability. FinAgent proposes a Model Context Protocol (MCP) server that acts as a dynamic context firewall, offering fine-grained access control on a per-request basis for agent-visible data. The system inte- grates Graph-based Retrieval-Augmented Generation (GraphRAG) with vector retrieval to support multi-hop reasoning on financial documents and knowledge graphs. An orchestrator manages specialized agents in a planner-worker-reflect framework, ensuring auditable and explainable results. We showcase FinAgent’s effec- tiveness via a functional prototype that processes simulated financial data, grounds facts via graph traversal, and enforces strict privacy boundaries while providing actionable portfolio analysis.

Introduction

The text describes the growing need for intelligent financial AI systems due to increasingly complex markets and large volumes of data. However, current solutions face key challenges: protecting sensitive user data, providing transparent explanations, and performing advanced multi-step (multi-hop) reasoning across interconnected financial information.

Existing systems fall into two categories:

• Cloud-based systems offer strong capabilities but raise privacy concerns because users must share sensitive data.
• Local systems preserve privacy but lack advanced analytical and reasoning abilities.
  Additionally, many systems operate as “black boxes,” offering little transparency in decision-making.

To address these issues, the paper introduces FinAgent, a privacy-focused, explainable, and intelligent financial assistant. Its main innovations include:

• A Model Context Protocol (MCP) for fine-grained privacy control over what data AI agents can access
• A GraphRAG framework combining vector search and knowledge graphs for multi-hop financial reasoning
• A multi-agent orchestration system that plans, executes, and reflects on tasks for structured decision-making
• Built-in explainability tools that provide traceable reasoning and confidence scores

FinAgent uses a local-first architecture, ensuring data privacy while maintaining strong analytical capabilities. Its system consists of components like an MCP server (for privacy enforcement), an orchestrator (for planning tasks), specialized agents (for analysis, risk, memory, and explanation), a knowledge layer (vector and graph databases), and an audit store (for traceability).

The paper also positions FinAgent within related work, highlighting limitations of existing financial AI, privacy-preserving methods, retrieval-augmented generation (RAG), and multi-agent systems. FinAgent advances these areas by integrating privacy control, explainability, and graph-based reasoning into a unified system.

Conclusion

This work introduces FinAgent, a privacy-preserving multi-agent financial intelligence system that tackles core challenges in automated financial analysis. By incorporating the Model Context Protocol for fine- grained context control, integrating GraphRAG for multi-hop reasoning, and implementing explicit orchestration with comprehensive audit trails, FinAgent demonstrates that advanced financial AI can be built with strong privacy guarantees and explainability. Our evaluation shows that GraphRAG substantially outperforms vector-only retrieval on multi-hop queries (87% vs. 62% accuracy), MCP achieves per- fect privacy preservation against adversarial prompts (0% leakage), and explainability mechanisms receive high user ratings (4.2/5.0). The system delivers ac- tionable portfolio insights with full provenance track- ing in an average of 3.2 seconds per query. FinAgent’s architecture provides a blueprint for privacy-aware agentic systems in high-stakes do- mains. By separating context governance, reasoning, and explanation into distinct layers, the system pro- motes modularity, testability, and regulatory compli- ance. Although limitations remain in planning flex- ibility and knowledge graph coverage, the working prototype demonstrates technical feasibility and es- tablishes a foundation for production-grade financial AI assistants. As AI systems increasingly participate in con- sequential decision-making, FinAgent’s emphasis on privacy, explainability, and auditability addresses key requirements for responsible deployment. Future work will extend these principles to other high-stakes domains that demand trustworthy automated assistance.

References

[1] Dwork, C. and Roth, A. (2014). The algorithmic founda- tions of differential privacy. Foundations and Trends in Theoretical Computer Science, 9(3-4):211–407. [2] Edge, D., Trinh, H., Cheng, N., Bradley, J., Chao, A., Mody, A., Truitt, S., and Larson, J. (2024). From local to global: A graph rag approach to query-focused sum- marization. arXiv preprint arXiv:2404.16130. [3] Gravitas, S. (2023). Autogpt: An autonomous gpt-4 exper- iment. [4] urlhttps://github.com/Significant-Gravitas/AutoGPT. Jiang, Z., Xu, D., and Liang, J. (2021). A deep reinforce- [5] ment learning framework for the financial portfolio [6] management problem. IEEE Transactions on Neural Networks and Learning Systems, 32(8):3590–3602. [7] Kairouz, P., McMahan, H. B., Avent, B., et al. (2021). Advances and open problems in federated learning. Foundations and Trends in Machine Learning, 14(1- 2):1–210. [8] Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Ku¨ttler, H., Lewis, M., Yih, W., Rockta¨schel, T., Riedel, S., and Kiela, [9] D. (2020). Retrieval-augmented generation for knowledge-intensive nlp tasks. In Advances in Neu- ral Information Processing Systems, volume 33, pages 9459–9474. [10] Schick, T., Dwivedi-Yu, J., Dess`?, R., Raileanu, R., Lomeli, M., Zettlemoyer, L., Cancedda, N., and Scialom, T. (2023). Toolformer: Language models can teach themselves to use tools. arXiv preprint arXiv:2302.04761. [11] Wooldridge, M. (2009). An Introduction to MultiAgent Sys- tems. John Wiley & Sons, 2nd edition. [12] Wu, S., Irsoy, O., Lu, S., Dabravolski, V., Dredze, M., Gehrmann, S., Kambadur, P., Rosenberg, D., and Mann, G. (2023). Bloomberggpt: A large language model for finance. In Proceedings of the 61st An- nual Meeting of the Association for Computational Linguistics, pages 5023–5043. [13] Yang, Y., Uy, M. C. S., and Huang, A. (2020). Finbert: A pretrained language model for financial communica- tions. arXiv preprint arXiv:2006.08097.

Copyright

Copyright © 2026 Dr. G. Murali, Tenali Sujith Kumar, Sk. Faisal Ahmed, U. Hemanth, P.V.S. Subhash. 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.