Show and tell: 7M tokens, 4 hours, $0 — real session proof

[KaiFelixBennett]

What I built

This repo is the exact setup I used for a 4-hour autonomous coding session with Hermes Agent on my local hardware:

• 7,256,671 tokens processed
• Claude Opus 4.7 equivalent cost: ~$94.34
• Actual cost: $0

Hermes planned and executed a multi-step code migration (iteration #7 of a background task), calling tools, editing files, and looping autonomously while I controlled it from Telegram.

The stack

• Hermes Agent — agentic loop, tools, Telegram gateway
• llama.cpp — local inference (AMD HIP backend, Qwen3.6-27B-MTP Q4_K_M)
• LiteLLM — Anthropic API bridge for Claude Code
• GPU: AMD Radeon AI PRO R9700, 32 GB VRAM

If you try it

Drop your hardware setup and token speeds here — especially interested in NVIDIA CUDA and Apple Silicon reports.

Setup: README

[KaiFelixBennett]

Cross-posted deeper benchmark findings + open questions here and I am collecting external critique in one place: https://www.linkedin.com/posts/kai-bennett-698732276_ai-opensource-developertools-activity-7465455904142864385-nbAJ. Add your take there or here; I will fold key insights into the docs.