Goose is an open-source AI agent framework from Aaif Goose, built with Rust and TypeScript. It functions as an AI agent or AI agent framework, connecting to various Large Language Model providers like OpenAI and Anthropic. Goose integrates with external tools and software through Model Context Protocol (MCP) servers, which act as extensions. It also operates as an Agent Client Protocol (ACP) server, making its AI agent capabilities available to compatible clients. Goose is designed for automating software development and other computer-based tasks, offering extensibility and security features.

What does Goose connect to?

Goose connects to multiple Large Language Model (LLM) providers, including OpenAI, Anthropic, Ollama, Google, and Amazon Bedrock. It integrates with external tools and software via Model Context Protocol (MCP) servers, which serve as extensions. Goose acts as an MCP host for these extensions and can also function as an Agent Client Protocol (ACP) server. This design allows it to expose core AI agent capabilities to compatible clients and interact with a broad ecosystem of tools.

Core features and tools

Goose exposes its AI agent capabilities as an Agent Client Protocol (ACP) server to clients like Zed, JetBrains, and VS Code. Callable tools are primarily exposed by MCP server extensions. Built-in extension categories include developer tools, file system operations, screen capture, and computer control. A specific example is the static_map_image_tool from the Mapbox MCP Server. Goose supports standard IO (stdio) and streamable HTTP for its MCP integrations, ensuring broad compatibility with various clients and protocols.

Who is it for?

Goose targets developers, AI engineers, and organizations needing an extensible AI agent for automating software development and other computer tasks. Its design supports complex workflows such as codebase management and debugging. The framework's ability to integrate with various LLM providers and external tools via MCP servers makes it suitable for environments requiring adaptable AI assistance. Skip this if your primary need is a simple, standalone AI tool without external integrations or advanced agent capabilities.

How to get started

Authentication for Goose primarily involves managing API keys for Large Language Model providers, allowing users to "bring their own keys". For specific MCP server extensions, setup often involves configuring access tokens or other credentials within Goose's configuration files, typically located at ~/.goose/config.json or ~/.config/goose/config.json on macOS/Linux, and %APPDATA%\Goose\config.json on Windows.

Use cases

• Codebase management, such as reviewing large pull requests, understanding complex changes, or migrating legacy codebases.
• Debugging, where Goose can guide a developer through intricate processes.
• Acting as a CI/CD bot, running in daemon mode with Agent Client Protocol (ACP) or orchestrated with scripts and recipes within a CI job runner.

Goose is an open-source AI agent framework designed for developers and organizations. It connects to major LLM providers and uses MCP servers for tool extensions, offering capabilities like codebase management and debugging. Security features include prompt injection detection and tool permission controls. Goose supports integration with popular development environments and CI/CD pipelines.