Local Agentic Infrastructure Workspace

This repository provides a fully localized, declarative development and execution environment for agentic infrastructure automation. It tightly integrates Loki (the agent execution engine), a local Ollama instance leveraging hardware acceleration, and Rescile-CE (Community Edition) via the Model Context Protocol (MCP).

The entire workspace is managed declaratively using NixOS / Nix Flakes and orchestrated seamlessly via direnv, ensuring reproducible dependencies across identical environments without polluting global system paths.

Architecture Overview

The workspace establishes a secure, local loop for translates high-level infrastructure requirements into deterministic execution sequences (Python, Pulumi, Terraform, Ansible, and Zsh).

• Loki (Agent Core): Processes localized reasoning prompts and orchestrates tool calling.
• Ollama (Local LLM Server): Handles on-device inference using optimized models (e.g., qwen2.5-coder, gemma3) accelerated directly via Intel GPU runtimes (xe kernel module / Level Zero).
• Rescile-CE: Acts as the NextGen Automation Server, exposing environment dependencies, multi-cloud topologies, and state validation to the agent via an HTTP MCP endpoint.

---

Prerequisites

To utilize this workspace, the following components must be installed on the host system:

• Nix Package Manager with Flakes enabled.
• direnv integrated with the host shell (zsh, bash, etc.).

---

Quickstart

1. Initialize the Environment

Navigate into the repository directory. direnv will automatically evaluate the flake.nix, build the specific version of Loki from source, inject all necessary platform tools (Ansible, Terraform, Pulumi), and configure the environment variables:

cd local_agents
direnv allow

2. Launch the Local Model Server

Due to isolation constraints within direnv and GPU driver bindings, the local Ollama daemon is decoupled from the automated shell hook to prevent terminal blocking. Start the model server in a separate terminal split or window within this directory:

start-ollama

3. Synchronize Models

In the terminal session running the workspace environment, download the preferred model into the localized workspace storage (.ollama/models):

ollama pull qwen2.5-coder:7b

4. Run the Agent

Ensure that the Rescile-CE server is active and listening on port 7600. Then, initialize the agent interface:

loki

---

Directory Structure & Configurations

The flake enforces a strict local state layout to keep the host system clean and comply with strict data sovereignty standards:

• flake.nix: Defines the deterministic build for loki-ai, injects the Intel Compute Runtime, and structures environment exports.
• .ollama/: Localized directory holding all downloaded LLM weights. Excluded from version control via .gitignore.
• ~/.config/loki/config.yaml: Automatically generated on shell initialization. Configures the OpenAI-compatible client endpoint pointing to the local Ollama instance.
• ~/.config/loki/functions/mcp.json: Declaratively provisioned to bind the rescile-automation MCP HTTP server into Loki's capabilities.

Links

• based on https://github.com/Dark-Alex-17/loki