Authority AI Editorial
Goose: Block's extensible local agent
Goose is a general-purpose AI agent that runs on your own machine, as a native desktop app, a command-line tool, or an embeddable API, rather than as a hosted service you talk to over the web. It was built by Block and, since December 2025, has been a founding project of the Agentic AI Foundation (AAIF) at the Linux Foundation, alongside the Model Context Protocol and OpenAI's AGENTS.md standard. Anyone who wants one agent that can move between coding, research, writing, and general automation without switching tools is the intended user.
What it does
Goose handles code, workflows, automation, research, and data analysis inside a single agent loop that runs locally. It connects to over 70 extensions through the Model Context Protocol, the open standard for linking a model to tools such as databases, browsers, GitHub, and Google Drive, so adding a capability is usually a matter of enabling an existing extension rather than writing new integration code. It also supports subagents, independent agents Goose can spin up to handle a parallel piece of work such as a code review or a research pass, which keeps the main conversation from filling up with intermediate steps. Recipes are portable YAML files that capture a workflow, including its parameters and any subrecipes, so a sequence a person builds once can be shared with a team or run unattended in a CI/CD pipeline.
MCP Apps extend this further by letting an extension render an interactive UI component directly inside Goose Desktop, rather than replying with text alone. That matters for tasks where a chart, a form, or a status panel communicates the result better than a paragraph would, and it means the extension author decides how the output looks rather than leaving every result as plain text in a chat window.
How you use it
Installation is a download of the desktop app for macOS, Linux, or Windows, or a CLI install via a bash script, and Goose can be pointed at more than 15 LLM providers, including Anthropic, OpenAI, Google, Ollama, Azure, Bedrock, and OpenRouter. It also supports using an existing Claude, ChatGPT, or Gemini subscription instead of a separate API key, which matters for anyone who does not want to manage another billed account. Day to day, a user opens a session, gives Goose a task in plain language, and enables whichever MCP extensions the task needs; recipes let a repeated task, such as a weekly report or a repo audit, be saved and rerun without re-explaining it each time. Teams can also build their own "goose distro," a custom distribution preconfigured with specific providers, extensions, and branding for internal rollout.
What to know before adopting it
Goose is written in Rust and released under the Apache 2.0 license. Its GitHub repository moved from block/goose to aaif-goose/goose in April 2026 as part of the transfer to the Agentic AI Foundation; the maintainers describe the project itself, its community, and its development pace as unchanged by the move, and old links redirect automatically. Security features are built into the agent loop rather than bolted on afterward: prompt injection detection, per-tool permissions, a sandbox mode, and what the documentation calls an adversary reviewer. Because it runs locally and is provider-agnostic, cost is whatever the chosen model provider charges, or nothing extra if a user brings an existing chat subscription.
Governance is worth a separate note for anyone evaluating long-term dependency risk. Goose is no longer a single company's internal tool released as open source on the side; it is a founding project of a Linux Foundation body with platinum members including AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI. That spread of backers is a different risk profile than a project maintained by one team inside one company, for better or worse depending on what a team weighs most.
Where it sits next to similar tools
Goose's closest comparison point is any agent framework built around MCP as the extension layer, since that is now a shared standard rather than a Goose-specific choice. What differentiates it is the combination of a real desktop app, a CLI, and an embeddable API around the same agent core, plus subagents and recipes as first-class features rather than something a team has to script on top. For anyone working on the instruction side of a Goose or similar setup rather than the extension side, the AI Ultimate Skill Collection offers 3,029 ready-made skills in 12 packs, written procedures your agent follows step by step, which covers the instruction layer a local agent still needs on top of its tool connections.