Skip to content

AI Tools

Jump to bottom
efrasoft edited this page Jul 2, 2026 · 2 revisions

AI Tools

HarnessOS ships with a curated set of AI coding tools pre-configured and ready to use.

The local AI brain (v2026.07.02.1)

Since v2026.07.02.1, Ollama is wired into the whole system as a single local brain: one model-routing config, one system-context server, every tool connected. Everything runs locally — no API key, no cost, nothing leaves your machine.

Shell AI — wtf, ask, Ctrl+X Ctrl+A

Command What it does
wtf Diagnoses the last failed command — sends the exact command, its exit code and live system context to the local model
ask "question" One-shot answer with real system context (kernel, services, disk, recent commands)
Ctrl+X Ctrl+A Explains the command currently typed at the prompt (risks included) without losing your input

When a command fails with a "real" error (exit code > 1), the shell prints a hint:

❯ systemctl statuss ollama
Unknown command verb 'statuss', did you mean 'status'?
↯ exit 2 — escribe 'wtf' para diagnóstico con IA
❯ wtf

If the answer contains commands, you get an approval menu ([1] [2] a=all n=skip) before anything runs.

Hardware-tuned models — harness-tune-ai

On first boot, harness-tune-ai detects GPU VRAM and system RAM and writes the optimal model tier to ~/.config/harness/config.toml:

Tier Hardware Example models
high ≥ 20 GB VRAM qwen3:27b, deepseek-r1:14b
mid ≥ 10 GB VRAM qwen2.5-coder:14b
base ≥ 5 GB VRAM / ≥ 14 GB RAM qwen2.5-coder:7b, llama3.2
low anything else qwen2.5-coder:3b, llama3.2

Machines without a dedicated GPU get a +cpu variant: interactive diagnosis modes (doctor, logs) use fast non-reasoning models — a chain-of-thought model on CPU takes 6+ minutes per answer. The chosen models are pre-pulled in the background so the first wtf doesn't stall on a download. Re-run after a hardware change: harness-tune-ai --force.

System MCP server — harness-mcp

/usr/local/bin/harness-mcp is a zero-dependency MCP (Model Context Protocol) server that exposes read-only system tools to any AI agent:

Tool What the agent sees
system_overview kernel, uptime, memory, disk, failed units, GPU
journal systemd journal (filter by unit, errors only, current boot)
service_status full systemctl status of one unit
pacman package info, file ownership, orphans, pending updates
docker_ps containers and their state
ollama_models local models + the routing config

Claude Code gets it registered automatically on first boot (claude mcp add --scope user harness); OpenCode loads it from ~/.config/opencode/opencode.json. Any other MCP client can use it: command = /usr/local/bin/harness-mcp. Since every tool is read-only, agents can see the machine but any mutation still goes through your normal command approval.

Editors, preconfigured

  • OpenCode (opencode) — ships with the local Ollama provider (llama3.2, qwen2.5-coder, deepseek-r1) and harness-mcp already configured. Just cd into a project and run opencode, then pick a model with /models.
  • VS Code + Continue — the Continue extension (installed from Open VSX on first boot) auto-detects every local Ollama model for chat, edit and autocomplete.

Claude CLI

Claude Code is Anthropic's official CLI for the Claude AI assistant. It provides an interactive terminal interface for AI-assisted coding.

Install (auto-installed on first boot with internet):

npm install -g @anthropic-ai/claude-code

Launch:

claude
# or from Hyprland: SUPER + C

First run: Claude will prompt you to log in with your Anthropic account at claude.ai.

Common usage:

claude                    # Interactive session
claude "explain this"     # One-shot query
claude --help             # All options

Ollama (Local LLMs)

Ollama runs large language models locally — no internet required after downloading.

Service: ollama.service is enabled and starts automatically.

Pull a model (first run, requires internet):

ollama pull llama3.2       # ~2 GB, recommended for most hardware
ollama pull codellama      # Specialized for code generation
ollama pull deepseek-coder # Strong code model

Run a model:

ollama run llama3.2
# or from Hyprland: SUPER + O

List installed models:

ollama list

API: Ollama exposes a REST API at http://localhost:11434. Compatible with OpenAI API format.

curl http://localhost:11434/api/generate -d '{
  "model": "llama3.2",
  "prompt": "Write a Python function to sort a list"
}'

GitHub Copilot CLI

Copilot in the terminal via the gh CLI extension.

Install (auto-installed on first boot):

gh extension install github/gh-copilot

Authenticate:

gh auth login

Usage:

gh copilot suggest "list all docker containers sorted by size"
gh copilot explain "$(cat script.sh)"

Docker with GPU support

Docker is pre-configured with NVIDIA container runtime support (when NVIDIA GPU is detected by the installer).

# Run a container with GPU access
docker run --rm --gpus all nvidia/cuda:12.0-base nvidia-smi

# Run Ollama in Docker with GPU
docker run -d --gpus all -p 11434:11434 ollama/ollama

Recommended workflow

┌─────────────────────────────────────────────┐
│  Your project directory                      │
│                                              │
│  $ claude          ← AI pair programmer     │
│  $ ollama run ...  ← Local reasoning        │
│  $ code .          ← VS Code editor         │
│  $ git commit      ← Version control        │
└─────────────────────────────────────────────┘

Clone this wiki locally