ai-tools

Scripts for running local AI tools across multiple machines.

Structure

Path	Contents
`ik-llama/`	CPU-only LLM inference with ik_llama.cpp — models, server, OpenCode config (ProBook, i9)
`ollama/`	GPU-accelerated Ollama inference for Apple Silicon (MacBook Air M4)
`tools/`	Install/update scripts for AI coding tools (Claude Code, OpenCode, Codex, nvm, hf, RTK, context-mode, claude-mem, Graphify)
`docker/openwebui/`	Open WebUI docker-compose for Ollama

tools/

Install and update scripts for AI coding tools. Core CLIs are installed via Homebrew where available; Node-based agent integrations use npm/npx. Run scripts individually or use update-all.sh to update everything at once.

Machine detection

Scripts detect the i9 work PC via proxy environment variables (tools/_brew-i9.sh). On i9, brew runs via sudo -n -u brewuser. On all other machines, brew runs directly. To enable on i9: add the brew binary to sudoers with the NOPASSWD flag.

update-all.sh

./update-all.sh

Order: nvm-update.sh → claude-update.sh → opencode-update.sh → codex-update.sh → hf-update.sh → hermes-update.sh → rtk-update.sh → context-mode-update.sh → claude-mem-update.sh → graphify-update.sh → ollama-update.sh → ollama-models-update.sh (ollama only runs if installed) → brew upgrade (all packages) → brew cleanup --prune=all. Each *-update.sh upgrades only if already installed and skips otherwise — run *-install.sh for new tools.

Claude Code

Script	What it does
`claude-install.sh`	`$BREW install --cask claude-code`
`claude-update.sh`	`$BREW upgrade --cask claude-code` (skips if not installed)

OpenCode

Script	What it does
`opencode-install.sh`	`$BREW install opencode`
`opencode-update.sh`	`$BREW upgrade opencode` (skips if not installed)

Codex (OpenAI CLI)

Script	What it does
`codex-install.sh`	`$BREW install --cask codex`
`codex-update.sh`	`$BREW upgrade --cask codex` (skips if not installed)

nvm / Node

Skipped entirely on i9 — work provides its own Node stack.

Script	What it does
`nvm-install.sh`	Installs nvm via brew, adds shell config block, installs LTS Node
`nvm-update.sh`	`brew upgrade nvm`, reinstalls LTS migrating global packages, `npm update -g`

Shell config detection: zsh → ~/.zshrc, bash+macOS → ~/.bash_profile, bash+Linux → ~/.bashrc.

NVM block appended (idempotent):

export NVM_DIR="$HOME/.nvm"
[ -s "$(brew --prefix)/opt/nvm/nvm.sh" ] && . "$(brew --prefix)/opt/nvm/nvm.sh"

Hugging Face CLI (hf)

Script	What it does
`hf-install.sh`	`$BREW install hf`
`hf-update.sh`	`$BREW upgrade hf` (skips if not installed)

The binary is called hf (not huggingface-cli). Installed via brew formula hf.

Hermes Agent (Nous Research)

Script	What it does
`hermes-install.sh`	`$BREW install hermes-agent`
`hermes-update.sh`	`$BREW upgrade hermes-agent` (skips if not installed)

RTK (Rust Token Killer)

Script	What it does
`rtk-install.sh`	`$BREW install rtk`, then initializes Claude Code, Codex, and OpenCode integrations
`rtk-update.sh`	`$BREW upgrade rtk`, then refreshes Claude Code, Codex, and OpenCode integrations (skips if not installed)
`rtk-init.sh`	Runs `rtk init -g`, `rtk init -g --codex`, and `rtk init -g --opencode`

RTK compresses command output before it reaches coding agents. After install/update, restart the affected agent sessions so hooks/plugins are loaded.

Context Mode

Script	What it does
`context-mode-install.sh`	`npm install -g context-mode`, then configures Claude Code, Codex, and OpenCode
`context-mode-update.sh`	`npm update -g context-mode`, then refreshes integrations (skips if not installed)
`context-mode-init.sh`	Re-applies integrations without reinstalling the npm package
`opencode-local-speed.sh`	Disables OpenCode-side helpers that can add latency around local model writes

Default integrations:

Codex: registers context-mode as a user MCP server.
OpenCode: skipped by default for local-model speed. Set CONTEXT_MODE_ENABLE_OPENCODE=1 to write the context-mode local MCP entry into ~/.config/opencode/opencode.json.

Claude Code plugin installation is opt-in because it runs a background Bun plugin process. Enable it explicitly with:

CONTEXT_MODE_ENABLE_CLAUDE=1 ./context-mode-install.sh
CONTEXT_MODE_ENABLE_OPENCODE=1 ./context-mode-init.sh

For local ik_llama.cpp/OpenCode sessions, if "Preparing write" is slow while CPU usage is low, disable OpenCode-side helpers and compare:

./opencode-local-speed.sh status
./opencode-local-speed.sh fast

fast disables context-mode in opencode.json, backs up and clears configured plugin entries, and moves every auto-loaded file from ~/.config/opencode/plugins/ to ~/.config/opencode/plugins.disabled/. Restore the MCP/plugin files with:

./opencode-local-speed.sh restore

Restart the affected agent sessions after install/update.

Claude-Mem

Script	What it does
`claude-mem-install.sh`	Installs/enables claude-mem integrations without auto-starting the worker; skips Codex/OpenCode reinstall unless forced
`claude-mem-update.sh`	Reports the available version and skips reinstall by default, because upstream update/install is interactive
`claude-mem-kill.sh`	Stops the claude-mem worker and kills common orphaned Bun/Chroma/MCP subprocesses

Claude-Mem stores settings under ~/.claude-mem/settings.json and may install worker dependencies. Its worker is a long-running Bun service on port 37700 by default, with Chroma MCP/Python subprocesses for semantic search and an internal Claude Haiku process for memory compression when the Claude provider is used.

Worker commands:

npx -y claude-mem@latest start
npx -y claude-mem@latest status
npx -y claude-mem@latest stop
./claude-mem-kill.sh
curl http://localhost:37700/api/health

Claude Code plugin installation is opt-in because it starts extra local services, including Bun, Chroma MCP, and an internal Claude process. Enable it explicitly with:

CLAUDE_MEM_ENABLE_CLAUDE=1 ./claude-mem-install.sh

OpenCode is installed non-interactively. Codex CLI gets the marketplace and hooks registered non-interactively, but the plugin itself may still need to be selected once from Codex's /plugins UI. Restart Codex and OpenCode after install/update so hooks/plugins are loaded.

Operational findings:

Do not run claude-mem reinstall from update-all.sh: upstream claude-mem update currently maps to the same interactive install flow and can prompt with "Overwrite existing installation?".
If Claude Code behaves oddly after testing the Claude plugin, disable it with claude plugin disable claude-mem, stop the worker with npx -y claude-mem@latest stop, and check for leftovers with ps -eo pid,ppid,etime,cmd | rg -i 'bun|claude-mem|chroma-mcp|claude-haiku'.
To keep the plugin installed/enabled but clear a stuck worker, run ./claude-mem-kill.sh.
To clear the worker and disable the Claude Code plugin in one step, run ./claude-mem-kill.sh --disable-claude.
If Chroma MCP remains after stopping the worker, kill the leftover chroma-mcp process before restarting Claude.
Keep Claude-Mem enabled first in OpenCode/Codex only; enable the Claude Code plugin only when explicitly testing whether the memory benefit outweighs the background-process cost.

To intentionally reinstall/repair claude-mem during updates:

CLAUDE_MEM_UPDATE_REINSTALL=1 ./claude-mem-update.sh

To force the upstream Codex/OpenCode installer again:

CLAUDE_MEM_FORCE_INSTALL=1 ./claude-mem-install.sh

Graphify

Script	What it does
`graphify-install.sh`	Installs the official PyPI package `graphifyy[openai,ollama,sql,pdf,office]` with `uv tool install`, falling back to `pipx`; then registers Claude Code, Codex, and OpenCode
`graphify-update.sh`	Upgrades Graphify with the same default extras when managed by `uv tool` or `pipx`, then refreshes integrations
`graphify-init.sh`	Re-registers Claude Code, Codex, and OpenCode integrations and enables Codex `multi_agent = true`

Graphify's package name is graphifyy but the CLI is graphify. The default extras are openai,ollama,sql,pdf,office, so the preferred installer is uv tool install 'graphifyy[openai,ollama,sql,pdf,office]' on both Linux and macOS. Override with GRAPHIFY_EXTRAS=...; use GRAPHIFY_EXTRAS= for the base package only. On macOS, install uv with Homebrew (brew install uv) if it is missing; the install script will also do this when Homebrew is available. pipx install 'graphifyy[...]' is kept as a fallback for Linux systems that already use pipx. On the i9/proxy environment, the scripts pass --system-certs to uv tool install so corporate CA certificates are honored.

After install/update, restart Claude Code, Codex, and OpenCode sessions. Use it inside a project with:

/graphify .

Codex uses $graphify instead of /graphify.

Graphify overlaps partly with context/memory tooling, but it is more of a project knowledge-graph generator than a background memory worker. Keep OpenCode local-speed testing in mind before enabling extra graph/query instructions in large local-model sessions.

Ollama

Script	What it does
`ollama-install.sh`	`$BREW install ollama` (skipped on i9 — CPU too slow)
`ollama-update.sh`	`$BREW upgrade ollama` (skipped on i9)
`ollama-models-update.sh`	Pulls latest version of every installed model (`ollama list \| xargs ollama pull`)

docker/openwebui/

Open WebUI connected to a local Ollama instance.

cd docker/openwebui
docker compose up -d

Listens on port 3000
Connects to Ollama at host.docker.internal:11434
OFFLINE_MODE=True — no external calls
Proxy env vars explicitly cleared so container bypasses any system proxy
Data volume: ~/docker/openwebui/data

ik-llama/

CPU-only local LLM inference using ik_llama.cpp on two machines (HP ProBook + i9-13900). Models: Qwen3.6, Gemma4, GLM-4.7-Flash.

See ik-llama/README.md for full details: hardware, models, flags, benchmark scripts, performance, and lessons learned.

Generated benchmark output goes into bench-results/ and is intentionally ignored by git. Python bytecode caches such as __pycache__/ and *.pyc are also ignored.

ollama/

GPU-accelerated local inference using Ollama on Apple Silicon (MacBook Air M4, 24 GB). Uses Metal GPU offload. Auto-discovers installed models for OpenCode config. See ollama/README.md for speed guidance and Ollama vs direct llama.cpp notes.

./tools/ollama-install.sh               # install Ollama binary
./ollama/download-models-mac.sh         # pull models (GGUF + Ollama tags)
./ollama/setup-opencode-mac.sh          # auto-generates provider config from ollama list

Models for 24 GB:

qwen36-27b-iq4nl — Qwen3.6 27B dense IQ4_NL (~14 GB), best quality coding model
qwen3:14b — Qwen3 14B dense (~9 GB), solid fallback
gemma3:12b — Gemma3 12B dense (~8 GB), fast alternative

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

ai-tools

Structure

tools/

Machine detection

update-all.sh

Claude Code

OpenCode

Codex (OpenAI CLI)

nvm / Node

Hugging Face CLI (hf)

Hermes Agent (Nous Research)

RTK (Rust Token Killer)

Context Mode

Claude-Mem

Graphify

Ollama

docker/openwebui/

ik-llama/

ollama/

About

Uh oh!

Releases

Packages

Uh oh!

Contributors

Uh oh!

Languages

Name		Name	Last commit message	Last commit date
Latest commit History 96 Commits
docker/openwebui		docker/openwebui
ik-llama		ik-llama
ollama		ollama
tools		tools
.gitignore		.gitignore
AGENTS.md		AGENTS.md
README.md		README.md

Folders and files

Latest commit

History

Repository files navigation

ai-tools

Structure

tools/

Machine detection

update-all.sh

Claude Code

OpenCode

Codex (OpenAI CLI)

nvm / Node

Hugging Face CLI (hf)

Hermes Agent (Nous Research)

RTK (Rust Token Killer)

Context Mode

Claude-Mem

Graphify

Ollama

docker/openwebui/

ik-llama/

ollama/

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages