Getting Started with Goose – Local AI Agent

Use at your own risk. All guides and scripts are provided for educational purposes only. Always review and understand any code before running it — especially with administrative privileges. Test in a safe environment before using in production. Your system, your responsibility.

Goose is an open-source AI agent developed by Block (the company behind Square and Cash App). Unlike a regular chat interface, Goose can actually do things — run shell commands, edit files, call external tools, and autonomously work through multi-step tasks. It connects to any LLM you choose, including local models via Ollama, which means your data can stay entirely on your own hardware.

This guide covers installing Goose on macOS and connecting it to a model provider. The follow-up guide covers using Goose to control remote machines via SSH.

What Makes Goose Different

Most AI assistants suggest what you should do. Goose just does it. It uses the Model Context Protocol (MCP) to connect to tools and extensions, which means it can interact with your filesystem, run commands, call APIs, and more — all autonomously, based on plain language instructions.

Goose is available as both a desktop app and a CLI tool. They share the same configuration, so you can switch between them freely.

Requirements

macOS (Apple Silicon or Intel)
Homebrew installed
An LLM provider — either:
- Ollama running locally (free, fully offline)
- An Anthropic API key (for Claude)
- Claude Code CLI installed (if you already have a Claude subscription)

Step 1 – Install Goose

Goose is available via Homebrew as a cask. Note that the cask name is block-goose, not goose — there’s a naming conflict with an unrelated database migration tool.

brew install --cask block-goose

Once installed, launch Goose from your Applications folder or Spotlight.

Step 2 – Choose a Model Provider

On first launch, Goose will ask you to configure a provider. You have several options:

Option A — Ollama (fully local, free)

If you have Ollama running locally, select Ollama as your provider and set the API host to:

http://localhost:11434

Then choose a model. For Goose to work well with tools and extensions, model choice matters a lot — more on that below.

Option B — Claude Code CLI

If you already have Claude Code installed and a Claude subscription, select Claude Code CLI and enter claude as the command. Goose will use your existing Claude session. This is the easiest path to reliable tool-calling without API costs beyond your subscription.

Option C — Anthropic API

Select Anthropic and enter your API key. Choose claude-haiku-3-5 for a cheap test, or claude-sonnet-4-5 for better quality. Claude handles tool-calling extremely well.

Step 3 – Configure Extensions

Extensions give Goose its superpowers. Go to Extensions in the sidebar. You’ll see two sections: Default Extensions (already available) and Available Extensions (off by default).

Recommended to enable:

Developer — runs shell commands, reads and writes files. Already on by default.
Computer Controller — lets Goose control macOS via clicks and keyboard input.
Memory — Goose learns your preferences over time.
Analyze — useful for reading logs, config files, and code.

You can toggle extensions on/off per session as well, which is useful when you want to limit what Goose has access to.

Step 4 – Start a Session

Click Start New Chat in the sidebar. Type instructions in plain language — Goose will figure out what tools to use.

Some examples to try:

Show me what's running on port 8080

Check disk usage on this machine and summarize it

List all running Docker containers

A Note on Model Quality and Tool-Calling

This is important and often glossed over: not all models handle tool-calling equally well.

Goose works by asking the model to decide when and how to call tools. Smaller or less capable models often generate the correct tool call syntax in their response but fail to actually execute it — they reason endlessly in text instead of acting.

From hands-on testing:

Model	Tool-calling reliability
Claude (any version)	Excellent — works first try
qwen2.5:32b	Good — reliable on most tasks
qwen2.5:14b	Inconsistent — works sometimes
gpt-oss:20b / Mistral variants	Poor — loops and hallucinates tool calls
Models under 14B	Generally not reliable for agentic use

If Goose seems to be thinking endlessly without doing anything, the model is likely the bottleneck. Switch to a larger or more capable model.

Hardware note: Running a 32B model locally requires at least 32GB of unified memory (Apple Silicon) with plenty of headroom for the OS and other apps. On a Mac Studio M1 Max with 32GB, a 14B model is the practical sweet spot for reliable agentic use.

Connecting to a Remote Ollama Instance

If you run Ollama on a more powerful machine (like a dedicated AI server with a discrete GPU), you can point Goose at it instead of running inference locally. In the provider settings, replace localhost with the IP of your Ollama server:

http://10.10.1.190:11434

This lets your Mac act as the agent frontend while a beefier machine handles inference — a useful setup if you have mixed hardware.