goose

goose is an open-source AI agent for general-purpose work: writing and testing code, researching, drafting text, automating workflows, or analysing data. It runs on your own machine and, according to the project, is built in Rust. The repository describes it as an extensible agent that goes beyond code suggestions: it installs, executes, edits, and tests, using whichever language model you choose. Anyone looking for an AI coding assistant that is not tied to a single model provider will find in goose an alternative to editor assistants with a hard-wired model.

How goose works

You use goose as a desktop app for macOS, Linux, and Windows or as a command-line tool; the documentation also lists an API. The agent receives a task, plans the steps, and calls tools for them, such as file access, shell commands, or external services. These tools arrive as extensions connected through the Model Context Protocol (MCP). The provider names more than 70 such extensions, including databases, APIs, browsers, GitHub, and Google Drive. That means many servers from the Skill Road catalog can be used directly in goose.

Model choice without lock-in

According to the documentation, goose works with more than 15 model providers, including Anthropic, OpenAI, Google, Ollama, OpenRouter, Azure, and AWS Bedrock. You can store your own API keys or keep using existing Claude, ChatGPT, or Gemini subscriptions. If you run a local model through Ollama, your requests stay inside your own network. This is the main difference from assistants that ship with a particular model and offer only that model's pricing and privacy terms.

Recipes, subagents, and MCP Apps

Recipes are portable YAML configurations that capture a recurring workflow together with its instructions and extensions, so you can share it with a team. Subagents let you hand sub-tasks to separate agents that work in parallel. MCP Apps add interactive interfaces that extensions can show inside the application. Together they let you grow goose from a one-off experiment into a repeatable working process.

Permissions and safety

You control how much freedom goose has through permission modes. According to the documentation there is fully autonomous operation (the default), manual approval of every tool use, a risk-based smart approval, and a chat-only mode with no file changes and no extensions. The mode can be switched during a session. The provider also names prompt-injection detection and a sandbox mode. Because the default is autonomous, it is worth deliberately choosing manual or smart approval for sensitive projects.

Privacy and telemetry

On first use, goose asks for permission to collect anonymous usage data; the choice can be changed at any time. According to the documentation it covers operating system, version, provider and model, how often tools are used, session metrics, and error types, but not conversations, code, tool arguments, or personal data. What goes to your chosen model provider is subject to that provider's own privacy terms.

License, operator, and cost

The source code is licensed under Apache 2.0. According to the README, the project belongs to the Agentic AI Foundation of the Linux Foundation. goose itself is free; costs arise only with the model providers you connect and follow their pricing pages.

Who goose fits

goose suits developers and technically minded teams who want an agent with their own model choice, local execution, and open source code, and who are willing to configure extensions and permissions themselves. If you would rather have a fully managed all-in-one tool with support and a fixed model, the catalog offers alternatives such as Cursor or Kiro. A sensible way to start is to install the desktop app, connect a model, enable a single extension in manual approval mode, and try a small task in a test project.