goose

goose is a general-purpose open-source AI agent that executes on your own machine rather than in a vendor's cloud. It began at Block, is written in Rust for speed and portability, and is now maintained by the Agentic AI Foundation under the Linux Foundation, licensed Apache 2.0.

Three ways to run it. A native desktop application for macOS, Linux and Windows; a full command-line interface for scripting and CI; and an API for embedding the agent in other systems. A companion iOS app lets you watch running agents and connect to as many remote goose instances as you have working for you. All three surfaces drive the same agent, so a session started in the terminal can be inspected from the desktop app.

Model-agnostic by design. goose works with more than 15 providers, including Anthropic, OpenAI, Google, Azure, Amazon Bedrock, OpenRouter and locally served models through Ollama. Support for the Agent Client Protocol means an existing Claude or ChatGPT subscription can back the agent instead of pay-as-you-go API keys.

Extensions over plugins. Capability comes from the Model Context Protocol: goose ships access to 70+ MCP extensions covering databases, HTTP APIs, browser control, GitHub and Google Drive, and you can point it at any MCP server you run internally. Skills and custom extensions can be added without forking the project.

Who it suits. Developers and operators who want an agent that can actually touch their environment — edit files, run builds, query databases, drive a browser — while keeping the code, the model choice and the execution host under their own control. It is free in the sense that matters for open source: no licence fee, no hosted tier required, and you pay only for whichever model provider you point it at. The CLI installs with a single script from the project's GitHub releases.