Secure Minions protocol enables encrypted Ollama‑frontier model collaboration

TL;DR

Secure Minions adds end‑to‑end encryption to the Minions framework, allowing local Ollama models (e.g., Gemma 3 4B) to collaborate with frontier cloud models (e.g., GPT‑4o) without exposing any plaintext during transmission or remote inference.

What Secure Minions Adds

Secure Minions extends the original Minions protocol by encrypting the entire local‑remote communication channel, even from the cloud provider. The key innovation is the use of NVIDIA Hopper H100 GPUs’ confidential computing mode to create a secure enclave that performs decryption, inference, and re‑encryption inside hardware‑protected memory.

Core Security Mechanism

Key Exchange – The local device and the H100 GPU perform a mutual key exchange.
Remote Attestation – The GPU proves it is genuine and running in confidential mode, preventing spoofed hardware.
Secure Enclave – All memory and computation inside the GPU are encrypted; even root users on the host cannot read plaintext.
End‑to‑End Encryption – Local LLM messages are encrypted before transmission, decrypted only inside the enclave, processed by the cloud LLM, re‑encrypted, and sent back.

"No plaintext is exposed – during transmission or remote LLM inference." – Hazy Research announcement

Performance Impact

Tests with prompts up to ~8 k tokens and large models such as Qwen‑32B show less than 1 % added latency, demonstrating that strong confidentiality does not compromise responsiveness.

Cost and Privacy Benefits

Cost Savings – By sending only the minimal encrypted tokens needed for orchestration, the protocol retains the 5×‑30× cloud‑cost reduction reported for the original Minions system while achieving 98 % of frontier model accuracy.
Privacy Upside – Sensitive context never leaves the local device in plaintext, addressing the primary privacy concern of local‑first LLM setups.

Getting Started

# Clone the Minions repository
git clone https://github.com/HazyResearch/minions.git
cd minions

# Optional: create a virtual environment
python3 -m venv .venv
source .venv/bin/activate

# Install the package in editable mode
pip install -e .

# Pull a local Ollama model (e.g., Gemma 3 4B)
ollama pull gemma3:4b

Running the Secure Demo

The repository includes a Streamlit app that demonstrates both the original Minion protocol and the Secure Minion (MinionS) protocol.

streamlit run app.py

Configure the app:

Remote Provider: "Secure"
Secure Endpoint URL: http://20.57.33.122:5056
Local Client: Ollama with your chosen model.

Secure Minions demo screenshot

Example Python Usage

from minions.clients.secure import SecureClient
from minions.clients.ollama import OllamaClient
from minions.minion import Minion

remote_client = SecureClient(
    endpoint_url="http://20.57.33.122:5056",
    verify_attestation=True,
)

local_client = OllamaClient(model_name="gemma3:4b")

protocol = Minion(local_client=local_client, remote_client=remote_client)

task = "How many grand slams did he win"
context = """John Doe, a legendary tennis player, ..."""

output = protocol(
    task=task,
    doc_metadata="file",
    context=[context],
    max_rounds=5,
)
print(output)

Run the script with python example.py.