Product managers and policy analysts are increasingly facing a dilemma: how to stress-test complex proposals against diverse stakeholder perspectives without exposing sensitive internal data to public cloud models. While large language models are exceptional at simulating human reactions, the requirement to upload proprietary documents to external servers has historically created an insurmountable security barrier for enterprise adoption. A new open-source project, mirollama, is now addressing this by enabling multi-agent simulations to run entirely within air-gapped, local environments.

The Architecture of Local Multi-Agent Simulation

mirollama functions as a self-contained simulation engine designed to operate without any external API calls. Drawing inspiration from the mirofish project, it leverages Ollama as its primary inference engine to execute models locally. The system is built to ingest PDF, Markdown, and plain text files, automatically transforming them into a structured pipeline that includes ontology mapping, knowledge graphs, persona generation, and final simulation reporting.

The technical stack is optimized for local performance and ease of deployment. It utilizes Flask as the backend framework, paired with Vue 3 and Vite to provide a responsive, browser-based interface. To eliminate the friction of manual environment configuration, the project includes full support for Docker Compose. For scenarios requiring external context, the system integrates with SearXNG, a privacy-focused, self-hosted metasearch engine, ensuring that even web-augmented research remains within the user's controlled infrastructure.

Shifting from Cloud API Dependency to Local Sovereignty

Traditional simulation workflows rely on sending sensitive policy data or internal strategy documents to providers like OpenAI or Anthropic. This creates a persistent risk of data leakage. mirollama fundamentally changes this dynamic by keeping all computational processes on the user's local server, effectively granting organizations total data sovereignty. Unlike standard chatbot interfaces that provide linear responses, mirollama orchestrates multiple agents, each assigned specific personas, to simulate complex organizational conflicts or market reactions through iterative interaction.

The project is currently optimized for lightweight, high-performance models, including Google’s Gemma 4 and various open-source models in the 20B to 120B parameter range. Developers can access the source code via the official GitHub repository. To deploy the environment, users must have Ollama installed and run the following command within the project directory:

docker-compose up --build

Once the containers are active, users can access the dashboard through a local browser to upload documents and define agent parameters. The development team is actively seeking contributions to expand model compatibility and is prioritizing updates that maximize inference efficiency on consumer and enterprise-grade hardware.

As local LLM inference performance continues to improve, the reliance on external cloud APIs for sensitive internal tasks will likely diminish, leading to a rapid increase in the adoption of on-premises AI agent frameworks over the next six months.