Scripts for running AI coding tools and local models across a few machines. This repository is public.
| Path | Contents |
|---|---|
llm/ |
Local models: one model catalog (models.json), shared commands for the i9 and the Mac, PowerShell for the ProBook (probook/), machine-only scripts in i9/ and mac/, measurements in llm/docs/ |
tools/ |
Install/update scripts for the AI CLIs (Claude Code, OpenCode, goose, Codex, Docker Agent, Antigravity, Pi, nvm, hf, RTK, context-mode, claude-mem, Graphify, Repomix, ccusage, Obsidian, Ollama, llama-swap), update-all.sh, and the cloud-provider pickers |
AGENTS.md |
Rules for AI agents working in this repo |
| machine | engine | model | start |
|---|---|---|---|
| ProBook (Windows) | mainline llama.cpp | Qwen3.6-35B-A3B-IQ4_NL-bartowski-MTP-32K |
llm\probook\llama-swap-setup.ps1, then opencode or llm\probook\claude-local.ps1 |
| i9 (Debian) | ik_llama.cpp | Qwen3.6-35B-A3B-UD-Q6_K-unsloth-MTP-128K |
llm/llama-swap-setup.sh |
| MacBook Air M4 | Ollama | Qwen3.6-35B-A3B-UD-Q2_K_XL-unsloth-32K |
llm/llama-swap-setup.sh |
Every machine runs the same chain: agents → local-proxy.py (9081) → llama-swap (9082, web UI with stats at
http://127.0.0.1:9082/ui) → engine. Stop with llm\probook\stop-llm.ps1 / llm/stop.sh. Thinking is off
by default everywhere. Details, first setup and the bare-server path: llm/README.md.
The folders ik-llama/, ollama/ and mlx/ were merged into llm/ on 2026-10-28, and the model names
changed on 2026-10-29: llm/README.md maps old to new.
Open WebUI (a chat UI with logins) runs as an optional container next to the chain: llm/README.md, "Open WebUI".
Generated benchmark output (bench-results/) and Python caches are ignored by git.