Goose is an open-source agentic developer tool created by Block (the company behind Square and Cash App) and released under the Apache 2.0 license. It runs locally in the developer’s terminal, connects to any LLM provider through a configurable backend, and extends its capabilities through Model Context Protocol (MCP) servers and a built-in extension system.
Key capabilities
Extension and MCP ecosystem — Goose ships with a library of built-in extensions (computer use, GitHub, Jira, Notion, databases) and supports any MCP server. Teams can build custom extensions in Python or TypeScript to connect internal tools.
Multi-provider LLM support — Goose is model-agnostic by design. It routes requests to Anthropic Claude, OpenAI, Google Gemini, Ollama-hosted local models, and others through a configurable provider layer.
File system and shell access — The agent reads, creates, edits, and deletes files and runs shell commands. A computer-use extension allows it to control the desktop GUI when terminal access is insufficient.
Browser automation — Built-in browser extension lets Goose navigate web pages, extract content, and test running applications.
Local execution — All processing runs on the developer’s machine; no data is sent to Block’s infrastructure. API calls go directly to the chosen LLM provider.
Autonomy level
Level 3 (supervised agent): the developer sees proposed actions before they execute. Goose can be configured for more autonomous operation in trusted environments.
Strengths
- Apache 2.0 license; no usage restrictions
- Richest built-in extension library among open-source coding agents
- Model-agnostic — switch providers without changing workflows
- Computer-use extension handles GUI automation that shell-only tools cannot
Limitations
- Computer-use extension is experimental and platform-specific
- Larger install footprint than simpler CLIs like Aider
- Less adoption than Claude Code or Cursor, so fewer community examples