codename goose

goose ships in three shapes from the same project. There is a native desktop app for macOS, Linux and Windows, a full command line tool for terminal and CI workflows, and an API for embedding the agent in your own software. The whole thing is written in Rust, which keeps it fast and portable rather than a wrapped web app. You pick the surface that matches how you work, and the same sessions and settings carry across.

goose does not lock you to one model. It works with more than 15 providers, including Anthropic, OpenAI, Google, Azure, Bedrock and OpenRouter, plus local models through Ollama. Through the Agent Client Protocol it can also run on a Claude, ChatGPT or Gemini subscription you already pay for, so you do not always need a separate API key. In practice that means you can swap models when one handles a task better or when a cheaper option makes sense.

goose was one of the earliest adopters of the Model Context Protocol, the open standard for connecting agents to tools and data. That gives it one of the deepest extension libraries around, with more than 70 documented MCP extensions and a growing list. You can connect GitHub, Google Drive, a browser, a database or your IDE, and write your own when nothing off the shelf fits. Extensions turn goose from a chat window into something that can actually touch your tools.

A recipe captures a workflow as a portable YAML file. You describe the instructions, the extensions, the parameters and any subrecipes, then share it with your team or run it in CI. This is what makes goose repeatable instead of a one-off chat: the same recipe produces the same steps every time. Recipes are also a neat fit for onboarding, since a teammate can run the exact process you built.

goose can spawn independent subagents to handle tasks in parallel, such as a code review, a research pass or file processing. Each one works on its own slice while the main conversation stays clean. For bigger jobs this keeps the context tidy and lets different parts of a task move at the same time instead of in a long single thread.

Because the agent runs on your machine, goose builds in ways to keep it in bounds. It includes prompt injection detection, tool permission controls, a sandbox mode, and an adversary reviewer that watches for unsafe actions. You can also set access controls so the agent asks before it acts rather than working autonomously. Since the source is open under Apache 2.0, you can check how those protections actually work.