goose Review

goose is a local AI agent for desktop, CLI, and API use. It works with 15+ providers including Anthropic, OpenAI, Google, Ollama, OpenRouter, Azure, and Bedrock, and you can use API keys or existing Claude, ChatGPT, or Gemini subscriptions through the Agent Client Protocol. It connects to 70+ extensions through MCP, supports MCP Apps for interactive UIs inside the desktop app, offers Recipes as reusable YAML workflows, can run subagents for parallel tasks, and lists security controls such as prompt injection detection, tool permissions, and a sandbox mode. It can also act as an Agent Client Protocol server for editors such as Zed, JetBrains, and VS Code. The project is part of the Agentic AI Foundation (AAIF) at the Linux Foundation.

The quickstart offers goose Desktop as a zip on macOS and Windows and DEB, RPM, or Flatpak on Linux, or the CLI through a single curl command from the releases page (on Windows, run it in Git Bash, MSYS2, or PowerShell). If you see a PATH warning, add goose to your PATH before running goose configure, and if prompted during setup the docs suggest choosing not to store credentials in the keyring. To configure a provider on Desktop, the welcome screen offers Quick Setup with an API key, a ChatGPT subscription, Agent Router by Tetrate, OpenRouter, or other providers, and in the CLI you run goose configure, choose Configure Providers, pick a provider, enter its API key, and choose a model (GitHub Copilot uses an auth code and a browser sign-in instead of a key). Start a session with Home on Desktop or goose session in an empty directory in the CLI, and enable an extension such as Computer Controller (web scraping, file caching, and automations) from the Extensions sidebar or goose configure, then resume with goose session -r. The README lists custom distributions, diagnostics and reporting, and a known-issues page, and the repository has moved to the aaif-goose organization (current version v1.54.0). The documentation outline for permission modes, tool permissions, and sandbox mode did not include body text when we read it, so consult the docs before running goose unattended.

You chat with goose in the desktop app or terminal, pick a provider, enable extensions, and let it work, with recipes capturing workflows you repeat. In the quickstart's example, goose plans and then builds a browser-based tic-tac-toe game and uses Computer Controller to open it. Because extensions run through MCP, capabilities grow without changing the core. This section reflects the quickstart, README, and OpenAlternative; we did not run goose hands-on.

• Desktop app, CLI, and API
• 15+ providers, including use of existing subscriptions through ACP
• 70+ MCP extensions and MCP Apps for interactive UIs
• Recipes: reusable YAML workflows
• Subagents for parallel tasks
• Security controls: prompt injection detection, tool permissions, and a sandbox mode
• ACP server mode for Zed, JetBrains, and VS Code

• General-purpose beyond coding, with a clean extension model through MCP
• Neutral governance under the Linux Foundation's Agentic AI Foundation
• Rust implementation and Apache-2.0 license, with about 55,000 stars
• Can reuse existing model subscriptions, which can cut cost, and has a fast quickstart on every desktop OS

• Extension quality varies, and each MCP extension widens the agent's reach
• The permission modes, tool permissions, and sandbox documentation were not readable in the pages we fetched, so verify them before unattended use
• Behavior depends on the model and extensions you choose
• Rapid releases (v1.54.0) mean the docs and defaults move, and a keyring prompt during setup can confuse first-time users
• Less IDE-integrated than editor-native agents

goose is free under Apache-2.0. Costs are model usage through your own API keys or existing subscriptions, and the pages we reviewed list no paid goose plan.

Claude Code and Codex are vendor coding agents tied to their own models. goose is vendor-neutral and broader in scope, with a desktop app, a CLI, MCP-based extensions, and recipes, which makes it more of a general work agent than a pure coding tool. Compared with OpenCode, Pi, and Cline, it leans toward extensibility and non-coding tasks and has a notably neutral governance home.

• Developers and analysts who want a general-purpose local agent with many MCP tools
• Teams that want to standardize workflows as reusable recipes
• Users who want to reuse existing model subscriptions through ACP

• People who want a purely editor-embedded assistant
• Teams that need a documented, centralized policy console
• Anyone uncomfortable granting an agent many tool permissions