PrivAgents: Agents and MCP servers for Confidential AI Workflows

Name: PrivAgents: Agents and MCP servers for Confidential AI Workflows
Author: UvrajSB

@UvrajSB

About PrivAgents: Agents and MCP servers for Confidential AI Workflows

PrivAgents is a privacy-first framework where intelligent agents interact with a secure MCP (Model Computation & Processing) server to perform encrypted data analysis using homomorphic encryption.

#mathgpt #math-solver #math-assistant

Basic information

Config

Add this server to your MCP-compatible client using the configuration below.

{
  "mcpServers": {
    "PrivAgent": {
      "command": "python",
      "args": [
        "-m",
        "venv",
        "venv"
      ]
    }
  }
}

Tools

No tools detected

We auto-extract tools from the README. The maintainer can list them under a ## Tools heading to populate this section.

Overview

What is PrivAgents?

PrivAgents is a privacy-first framework that enables autonomous agents (on-device or in the cloud) to perform secure data processing through trustless MCP servers using homomorphic encryption. It is designed for sensitive environments such as healthcare, personalization, and finance, where data must remain confidential during processing.

How to use PrivAgents?

Clone the repository, create a Python virtual environment, install dependencies, and start the MCP server from the root directory with PYTHONPATH=. python mcp/server.py. Then run either the Ollama agent (PYTHONPATH=. python agent/ollama_mcp_agent.py) for local execution or the OpenAI agent (PYTHONPATH=. python agent/openai_mcp_agent.py) for cloud execution.

Key features of PrivAgents

Supports both on-device (Ollama) and cloud-based (OpenAI) agents
Homomorphic encryption (BFV scheme) for end-to-end data confidentiality
Modular MCP server performs encrypted similarity calculations
Encrypted pipeline: data encrypted on device, processed, and decrypted locally
Demonstrated with movie preference matching using encrypted dot products

Use cases of PrivAgents

Privacy-preserving recommendation systems that never expose user preferences
Secure data processing in healthcare, finance, and personalization
Encrypted similarity matching for any domain (movies, music, shopping)
Educational demonstrations of homomorphic encryption in practical workflows

FAQ from PrivAgents

What encryption scheme does PrivAgents use?

PrivAgents uses the BFV homomorphic encryption scheme via TenSEAL to encrypt user data before sending it to the MCP server.

Which agents can I use with PrivAgents?

Two agents are supported: the Ollama Agent for on-device reasoning with local models, and the OpenAI Agent for cloud-based intelligence via API.

Where does the actual data processing happen?

The user’s input is encrypted on the device, sent to the MCP server for encrypted similarity calculations, and the results are returned in encrypted form to be decrypted locally on the device.

What dependencies are required to run PrivAgents?

The project requires Python, a virtual environment, and packages listed in requirements.txt (including TenSEAL). No additional runtime or cloud backend is required beyond the chosen agent’s API key.

Does the MCP server ever see unencrypted data?

No. The server only receives encrypted tensors and computes dot products under encryption. User data remains confidential throughout the entire workflow.