Hermes Claude Code Local

Hermes Local Stack — Local Agentic Coding with Claude Code, llama.cpp, Zero API Costs
A 4-hour autonomous coding session — 7,256,671 tokens — would have cost $94.34 on Claude Opus 4.7. With this stack: $0.

This repo wires Hermes Agent directly to llama.cpp and optionally bridges Claude Code through a local LiteLLM proxy — so your agent calls tools, edits files, and runs task loops 100% on your own hardware.
Which local model should you run? benchmark.securesight.ai compares measured agentic coding runs on AMD hardware, among them the Qwen3.6-27B used here and its successor Qwen3.8-27B: speed, quality of the result, raw logs. Most of the games the models built are playable in the browser. Data and code: local-ai-amd-benchmark.

  Real session  ·  4 hours  ·  7,256,671 tokens
  Same session with Claude Opus 4.7 would have cost ~$94.34  ·  Cost with this stack: less than a coffee

Session title: "Claude Code Migration im Hintergrund läuft #7" — Hermes autonomously ran a multi-step code migration in the background, 7th iteration, while the phone stayed in my pocket.

Quick Start

Windows (PowerShell)

Linux / macOS

That's it. On Windows the setup configures Hermes with llama.cpp and launches everything. On Linux and macOS it clones this repo to ~/hermes-claude-code-local, installs Hermes Agent and llama.cpp (Metal on Apple Silicon, Vulkan or CPU on Linux), downloads the model if you have none and points Hermes at the local server. Then start it:

What You Get

- Agentic coding, fully local — Hermes plans multi-step tasks, calls tools (web search, file edits, browser), and executes autonomously without cloud round-trips
- Hermes Agent running locally with llama.cpp as the inference engine
- Claude Code bridge via LiteLLM (optional) — Claude…