Back to system overview

server

Local inference flexibility. The RPC distributed model server handles requests seamlessly when the router tier is set to local.

Backend Agnostic

The hirn server doesn't force you into a single execution engine. It provides a unified RPC interface that seamlessly connects to multiple powerful backends:

  • llama.cpp: For highly optimized CPU/GPU offloading on consumer hardware.
  • vLLM: For high-throughput serving on larger VRAM setups.
  • exo / cactus: For experimental and specialized local routing architectures.

Distributed VRAM Management

Running large models often requires more VRAM than a single machine holds. The server intelligently pools and manages VRAM across multiple GPUs or even multiple devices on your local network, treating your home lab as a single unified inference cluster.

Zero Setup

Starts alongside your agent automatically. No complicated Docker configurations or Python environments required for standard operation.

Cross-Platform

Natively optimized for Windows, Linux, and macOS, providing bare-metal performance regardless of your OS setup.

Auto-Scaling

Dynamically unloads idle models from VRAM to free up system memory, only loading them when a router request comes through.