The Netyeti's Journal

AI Sysadmin Stack - Usage Guide

7 min read

AI Sysadmin Stack — Usage Guide

The full local AI toolkit for system administration: terminal AI, shell integration, log analysis, script generation, and remote SSH management — all backed by local Ollama models.

For installation instructions (fresh machine or reinstall), see 2026-04-01-AI-Sysadmin-Stack-Installation. For the GPU/llama.cpp setup and backstory, see 2026-03-22 and 2026-03-29-start.

What’s in This Stack

ToolBinaryPurpose
aichataichatPrimary terminal AI — roles, sessions, pipes
llmllmUnix-pipe AI, great for log/file analysis
aideraiderCode editing with AI (now Ollama-backed)
Python FabricfabSSH automation for remote hosts
fabaifabaiAI-planned remote management via SSH

All of these run against local Ollama models by default. No tokens sent anywhere.

Models and When to Use Each

ModelBest forSpeed (CPU)
llama3.1:8bGeneral sysadmin Q&A, reasoning, troubleshooting~2 t/s
qwen2.5:7bGeneral purpose, concise answers, quick queries~2 t/s
qwen2.5-coder:7bShell scripts, Python, Ansible, Dockerfiles~2 t/s
nemotron-3-nano:4bFastest option, good for simple tasks~3 t/s
qwen2.5-coder:1.5b-baseAutocomplete, one-liners, ultra-fast~6 t/s

Note on speed: Ollama has OLLAMA_VULKAN=true set in the systemd service as of April 2026. Whether this actually offloads to the Arc iGPU depends on Ollama’s version. If responses feel slow, use ai-gpu (below) which uses llama.cpp + Vulkan directly and is confirmed GPU-accelerated (28.8 t/s generation vs 2 t/s CPU). See 2026-03-29-start for measured benchmarks.

Shell Aliases Quick Reference

All of these are in ~/.zshrc. Run source ~/.zshrc or open a new terminal after any changes.

# aichat — the main tool
ai          # plain aichat, no role
aic         # sysadmin role (llama3.1:8b)
aid         # devops role — Ansible, Docker, K8s (qwen2.5-coder:7b)
ais         # script writing role (qwen2.5-coder:7b)
aix         # explain a command/concept (llama3.1:8b)
aif         # sysadmin role with persistent session (for multi-turn)
aifast      # no role, qwen2.5:7b (smaller/faster)
aicode      # no role, qwen2.5-coder:7b
 
# GPU-accelerated (requires llama-serve running — see below)
aigpu           # llama3.1:8b via GPU
aigpu-code      # qwen2.5-coder:7b via GPU, script role
aigpu-sys       # llama3.1:8b via GPU, sysadmin role
ai-gpu [model]  # start GPU server + open aichat in one command
 
# llm — pipe-friendly
lm          # llm with llama3.1:8b (default)
lmfast      # llm with qwen2.5:7b

Shell Functions

explain — understand any command

explain "find . -name '*.log' -mtime +7 -delete"
explain ls -lah /etc
explain "awk '{print $1}' access.log | sort | uniq -c | sort -rn"

ailog — analyse log output

# Pipe logs directly
journalctl -xe | ailog
tail -n 500 /var/log/nginx/error.log | ailog
 
# With a specific question
journalctl -u postgresql --since "1 hour ago" | ailog "focus on connection errors"
cat /var/log/syslog | ailog "any signs of disk failure?"

aigen — generate a command from a description

aigen "find all files modified in the last 24 hours larger than 100MB"
aigen "show all listening TCP ports with the process name"
aigen "rotate logs in /var/log/myapp older than 30 days"

aifix — diagnose a failed command

aifix "sudo systemctl restart nginx" "Job failed. See journalctl -xe for details."
aifix "ansible-playbook site.yml" "ERROR: The task 'Install packages' failed with: No package nginx found"

Keyboard Shortcuts (Shell)

KeyAction
Ctrl+KExplain the command currently typed in the shell buffer
Ctrl+GReplace typed description with a generated command

Example of Ctrl+G: Type show which processes are using the most memory and press Ctrl+G. The buffer is replaced with ps aux --sort=-%mem | head -20 (or similar). Review it, then press Enter to run.

aichat — Interactive Mode

aichat drops into a REPL when called without piped input:

aic                     # sysadmin role REPL
aic --session debug-1   # named session (saved, resumable)

Inside aichat:

# Change model mid-session
.model ollama:qwen2.5-coder:7b

# List available models
.models

# List roles
.roles

# Switch role
.role script

# Clear context
.clear

# Exit
.exit

Roles

RoleUse it for
sysadminLinux administration, services, networking, storage
devopsAnsible, Docker, docker-compose, Kubernetes, Talos
scriptWriting bash or Python scripts (outputs clean code blocks)
explainUnderstanding commands, man pages, log entries

Role files are in ~/.config/aichat/roles/ — edit them to tune behaviour.

GPU-Accelerated Path

For faster responses (confirmed ~3× speedup on the Arc iGPU), use the llama.cpp Vulkan server directly.

# Start the GPU server and open aichat (one command)
ai-gpu          # default: qwen2.5-coder:7b
ai-gpu llama    # llama3.1:8b
ai-gpu qwen     # qwen2.5:7b
ai-gpu nano     # nemotron-3-nano:4b (fastest)
 
# Or start the server separately and use GPU-backed aliases
llama-serve coder
aigpu-sys       # now talking to GPU backend

The GPU server runs at http://localhost:8081. aichat accesses it as the llama-gpu client. Models are listed as llama-gpu:qwen2.5-coder-7b.gguf etc.

Note: Only one model can be loaded in llama-server at a time. If you switch models, stop and restart with a different argument. Ollama (port 11434) is unaffected.

llm — Pipe Mode

Simon Willison’s llm tool is ideal for chaining into shell pipelines.

# Summarise anything
cat /etc/nginx/nginx.conf | lm "explain this config and flag any security issues"
df -h | lm "is anything close to full?"
ps aux | lm "what's consuming the most resources and is anything unusual?"
 
# Analyse a file
lm "what does this do?" < ~/bin/llama-serve
 
# With a specific model
echo "write a cron expression for every weekday at 3am" | llm -m qwen2.5-coder:7b
 
# Check available models
llm models list | grep Ollama

aider — Code Editing

aider is now configured to use qwen2.5-coder:7b via Ollama by default (~/.aider.conf.yml).

# Standard usage — Ollama backend, no cloud needed
aider src/app.py
 
# Using the GPU server (faster)
aider-local               # GPU server, coder model
aider-local llama         # GPU server, llama3.1:8b
 
# Override model directly
aider --model ollama/llama3.1:8b

Config lives at ~/.aider.conf.yml. Key settings:

  • Default model: ollama/qwen2.5-coder:7b
  • Editor model: same
  • Weak model (commit messages): ollama/qwen2.5-coder:1.5b-base
  • Auto-commits: off (review diffs manually)

fabai — AI-Assisted Remote Management

fabai generates shell commands for a task using Ollama, shows them to you, then optionally executes them over SSH via Fabric.

Basic usage

# Ask AI what to run on a remote host, then decide
fabai --host web01 "check disk usage and identify the top 5 largest directories"
fabai --host db01 --user ubuntu "is postgresql running and healthy"
fabai --host bastion "show failed SSH login attempts in the last hour"
 
# With a smaller/faster model
fabai --host web01 --model qwen2.5:7b "check memory usage"

Dry-run mode (generate and run locally)

fabai --dry-run "show all listening ports with process names"
fabai --dry-run "find log files larger than 500MB"

Interactive mode

fabai --interactive --host web01
fabai --interactive --host web01 --user ubuntu
# No host = runs locally
fabai --interactive

Prompts you for tasks in a loop. Type quit to exit.

Destructive operations

If fabai detects that a command is destructive (deletes files, restarts services, etc.), it warns you and requires explicit y confirmation before executing.

Python Fabric — Direct SSH Automation

For structured, repeatable automation, write a fabfile.py. A template is at ~/bin/fabfile-template.py.

# Copy the template to a project
cp ~/bin/fabfile-template.py ~/myproject/fabfile.py
 
# Run a task
source ~/.local/sysadmin-env/bin/activate   # activate the sysadmin venv
fab --host web01 disk-report
fab --host db01 log-check --lines=500
fab --host web01 sys-info
fab --host web01 service-health --service=nginx
 
# Multiple hosts
fab --host web01 --host web02 disk-report

The template’s tasks each collect data from the remote host and send it to Ollama for analysis. The ai-task task takes a natural language description:

fab --host web01 ai-task --describe="show failed logins in the last hour"

Using the sysadmin venv in scripts

For standalone Python scripts that use both Fabric and the Ollama library:

#!/home/netyeti/.local/sysadmin-env/bin/python3
from fabric import Connection
import ollama

Or activate first: source ~/.local/sysadmin-env/bin/activate

Keeping Models Warm

Ollama is configured with a 30-minute keepalive (OLLAMA_KEEP_ALIVE=30m). The first call after a cold start takes ~80 seconds to load the model into memory. After that, responses are immediate until the keepalive expires.

To pre-warm a model (e.g., in the morning or before a session):

echo "ready" | llm -m llama3.1:8b > /dev/null

To check what’s currently loaded:

curl -s http://localhost:11434/api/ps | python3 -m json.tool

Adding a New Model

# 1. Pull via Ollama
ollama pull <name>:<tag>
 
# 2. Add it to aichat config
#    Edit ~/.config/aichat/config.yaml — add under the ollama client's models list
 
# 3. Optionally symlink for llama-server
#    Get the blob hash from the manifest (see ~/bin/llama-serve or [[Ollama and My System]])
#    ln -sf /usr/share/ollama/.ollama/models/blobs/sha256-<hash> ~/models/<name>.gguf
#    Add it to the MODELS array in ~/.local/bin/llama-serve

Quick Diagnostics

# Is Ollama running?
systemctl status ollama --no-pager | head -5
 
# What's loaded?
curl -s http://localhost:11434/api/ps | python3 -m json.tool
 
# Is llama-server (GPU) running?
curl -s http://localhost:8081/health
 
# aichat models visible?
aichat --list-models
 
# llm default model?
llm models default
 
# Test everything quickly
echo "ping" | lm
echo "ping" | aichat -m ollama:qwen2.5:7b

Last updated: April 2026 — phoenix (ASUS Zenbook Duo UX8406MA, Ubuntu 24.04) For reinstallation steps, see 2026-04-01-AI-Sysadmin-Stack-Installation.

Graph View

Backlinks