goose-nest

Local AI workstation setup — Goose (Block's AI agent) + Ollama (local LLM inference) on NVIDIA GPU, managed with Ansible.

Prerequisites

• Ubuntu/Debian with NVIDIA GPU and drivers installed
• CUDA toolkit
• Ansible 2.18+
• gh CLI (for repo management)
• Classic sudo installed (package sudo, binary at /usr/bin/sudo.ws on Ubuntu). ansible.cfg pins become_exe to it because sudo-rs — now the default /usr/bin/sudo on recent Ubuntu — breaks Ansible's -K password handshake. If you're on a distro without sudo.ws, adjust become_exe in ansible.cfg to point at your classic sudo binary.

Quick Start

git clone https://github.com/eschoeller/goose-nest.git
cd goose-nest
ansible-playbook deploy.yml -K

You'll be prompted for sudo password for tasks that need root (Ollama install, systemd config, model directory creation).

Post-Install: Configure Goose

After the playbook completes, you need to configure Goose manually:

1. Select provider and model

goose configure

• Provider: Ollama
• Model: qwen3-coder:30b (recommended) or qwen2.5-coder:14b
• Telemetry: your choice

2. Enable extensions

By default, Goose has very few extensions enabled. For full agentic capabilities (file access, shell commands, code editing), enable these extensions via goose configure:

Extension	What it provides
developer	Shell access and code editing (essential)
code_execution	Token-efficient tool calls via code
computercontroller	Web scraping, file caching, automations
memory	Persistent memory across sessions
chatrecall	Search past conversations
autovisualiser	Data visualization and UI generation

Without the developer extension, Goose cannot read files or run commands — it will only be able to chat.

3. Verify

goose session

Ask Goose to read a file or run a command to confirm extensions are working.

Config location

Goose stores its configuration at ~/.config/goose/config.yaml.

What It Does

1. gpu_verify — Confirms NVIDIA GPU is present, reports VRAM/CUDA/driver info
2. ollama — Installs/upgrades Ollama, moves model storage to /games/models/ollama, configures systemd
3. models — Pulls configured LLM models
4. goose — Installs/upgrades Goose CLI

Both Ollama and Goose are automatically upgraded when new versions are available.

Configuration

Edit group_vars/all.yml to customize:

ollama_models_dir: /games/models/ollama   # Where models are stored on disk
ollama_host: "http://localhost:11434"

ollama_models:                             # Models to pull
  - qwen3-coder:30b                       # ~19GB - MoE, best tool-calling
  - qwen2.5-coder:14b                     # ~9GB  - Dense coding model
  - qwen2.5:7b                            # ~4.7GB - General purpose
  - deepseek-r1:7b                        # ~4.7GB - Reasoning

goose_default_model: qwen3-coder:30b

Running Individual Roles

Use tags to run specific parts:

ansible-playbook deploy.yml --tags gpu      # GPU check only
ansible-playbook deploy.yml --tags ollama    # Ollama setup only
ansible-playbook deploy.yml --tags models    # Pull models only
ansible-playbook deploy.yml --tags goose     # Goose setup only

Verification

ansible-playbook deploy.yml --check   # Dry run
ollama list                           # Verify models
goose -V                              # Verify Goose version
goose session                         # Start interactive session