goose

What it is

goose is an open source AI agent that runs on your machine. It comes as a native desktop app for macOS, Linux and Windows, as a CLI and as an API. You give it a task in plain words. goose makes a plan, then uses tools to carry it out. It reads and edits files, runs commands and calls other apps through MCP extensions. The README says it is for research, writing, automation and data analysis as well as code, and that it is part of the Agentic AI Foundation at the Linux Foundation. This listing holds Mindket’s setup notes for it, written from the maker’s README and docs.

1. You describe a task in plain words
2. goose plans it with the model you chose
3. Extensions run the tools: files, commands, other apps
4. You approve each step, or let goose run on its own
5. The session is saved, so you can come back to it

Who it suits

Developers who want an agent in the terminal or in a desktop app, working with the model they already pay for. It also suits people who write no code, since the README lists research, writing and data analysis. Teams can fork it into their own distribution with preset providers, extensions and branding.

If all you want is code suggestions in an editor, goose does more than you need.

Where to use it

goose’s docs cover each of these routes.

• Desktop app: macOS (download or Homebrew), Linux (DEB, RPM or Flatpak) and Windows (a zip you unpack)
• CLI: one install script for macOS, Linux and Windows in Git Bash or MSYS2, a PowerShell script on Windows, or WSL
• Claude Code or Codex as the model behind goose, through ACP, with your existing subscription
• Zed, as an ACP agent from its registry or with your own goose binary
• Your shell, with @goose at the prompt, and scripts or GitHub Actions through goose run
• Many model providers, and local models through Ollama or LM Studio

goose needs a model that can call tools. A model without tool calling can only chat, and every extension must be switched off for it.

How to set it up

Get it saves a zip of Mindket’s setup notes in your downloads folder, with the steps for each system from goose’s docs. The main routes are a desktop download or one line for the CLI.

1. Install the desktop app for your system, or run the CLI line below.
2. Choose how goose reaches a model. The desktop welcome screen offers a pasted key, a ChatGPT login, Tetrate or OpenRouter. In the CLI, goose configure lists the providers and asks for the key.
3. Start a session. In the desktop app, type in the box. In the CLI, run goose session in your project folder.
4. Before real work, pick the permission mode you want. The default is fully autonomous.

curl -fsSL https://github.com/aaif-goose/goose/releases/download/stable/download_cli.sh | bash
goose session

On a Mac, Homebrew installs the desktop app with brew install --cask block-goose and the CLI with brew install block-goose-cli. Add CONFIGURE=false before bash to skip the interactive setup.

How to use it

Talk to goose as you would to a developer. The Quickstart’s first prompt is below.

create an interactive browser-based tic-tac-toe game in javascript where a player competes against a bot

goose writes a plan, then the files. To let it open the game for you, enable the Computer Controller extension. In the desktop app, open Extensions in the sidebar and switch it on. In the CLI, run goose configure, choose Add Extension, then Built-In Extension, then Computer Controller. Resume the session with goose session -r and ask:

open the tic-tac-toe game in a browser

Example

You keep telling goose the same things about a project. You put a .goosehints file at the project root, with lines taken from the docs’ own example:

This is a simple example JavaScript web application that uses the Express.js framework.

Make sure to confirm all changes with me before applying.

Run tests with `npm run test` ideally after each change.

goose reads the file when a session starts and adds it to every request, so you stop repeating yourself.

The Developer extension must be on for hints to work. A file named AGENTS.md at the root is read too.

To make goose ask before it writes, switch the mode inside the session:

/mode approve

In the approval modes, the desktop app shows Allow and Deny buttons when goose wants to use a write tool, such as a file edit or a bash rm, cp or mv.

For newcomers and experts

If you are new. An agent differs from a chat box because it can act. goose reads your folder, edits files and runs commands, and it does so through extensions. goose runs in autonomous mode by default, and the Developer extension is on by default, so together they let goose change files and run commands without asking you. Start in a folder under version control. Stage your own work and leave goose’s edits unstaged until you have read them, as the Known Issues page advises. The Quickstart is a short tutorial: install goose, pick a model, build a small app and add an extension.

For experts. goose can use your Claude Code or Codex subscription as its model through ACP. You install an adapter, sign in with the CLI you already have and set one variable. goose then passes its extensions to that agent as MCP servers. Sessions on an ACP provider cannot be resumed or forked yet.

npm install -g @agentclientprotocol/claude-agent-acp
export GOOSE_PROVIDER=claude-acp

• Recipes: package extensions, prompts and settings into a file, then run it with goose run --recipe recipe.yaml
• Skills: a folder with a SKILL.md in ~/.agents/skills/ or in a project’s .agents/skills/, compatible with other agents that read Agent Skills
• Headless: goose run --output-format json for scripts and CI, with GOOSE_MODE=auto and a turn limit
• Zed: "args": ["acp"] in agent_servers runs goose as an agent inside the editor
• Distributions: preset providers, bundled extensions, branding and system prompts, under Apache 2.0

Good to know

goose acts on your machine, and a few of its paths reach out.

• Acts without asking: autonomous mode is the default, and with the Developer extension goose runs commands and changes files with no approval. The other modes are approve, smart_approve and chat
• Your model provider sees your work: your messages and what goose reads go to the provider you choose, under its retention and privacy policies. A local model keeps them on your machine
• Usage data: goose asks on first use, and the default is off. If you opt in, it counts things like your system, provider, model and tool names, and the docs say it leaves out conversations and code. Turn it off with GOOSE_TELEMETRY_ENABLED=false
• Keys: provider keys live in the system keyring, or in a plain text secrets.yaml where there is none. A key written in config.yaml is ignored
• Extension check: goose checks npx and uvx extensions against a malware database before it starts them. If that service is down, they still install
• Shell history: off by default. The flag --dangerous-include-history sends every command you type to your model provider, passwords included
• Removal: the docs list the folders to delete for each system, and a Keychain or Credential Manager entry to clear by hand
• License: Apache-2.0. The LICENSE appendix reads Copyright 2024 Block, Inc. The custom distributions guide asks forks not to use the Goose trademarks in ways that imply official endorsement