# Hermes Local Stack

This project wires the Hermes autonomous agent directly to a local llama.cpp model, and optionally bridges Claude Code through a local LiteLLM proxy, so an agent can plan tasks, call tools and edit files entirely on a user's own hardware with no per-token API cost.

Hermes talks directly to llama.cpp for its own planning and tool calls, which the project describes as fewer moving parts and simpler debugging, while a separate optional path lets Claude Code send its requests through a local LiteLLM proxy instead of Anthropic's API, since Claude Code expects Anthropic-style API behavior that Hermes does not need to match. Telegram integration lets a user control Hermes from another device, including with voice messages, while a task runs unattended.

The project ships a one-command setup script for Windows PowerShell or for Linux and macOS, uses multi-token prediction and speculative decoding for faster local inference with a specific quantized model, and includes self-healing behavior that restarts the LiteLLM proxy if it crashes.

It is packaged as Shell and PowerShell scripts around existing Hermes, llama.cpp and Claude Code installations.

Website: https://github.com/KaiFelixBennett/hermes-claude-code-local