Goose Review, Pricing & Alternatives (2026)

Goose is Block’s open source terminal AI agent for developers and technical teams. Launched on January 28, 2025, it began with software engineering work such as code migrations, dependency installation, test execution, and file changes, then expanded toward research, writing, automation, and data analysis. The project uses the Apache License 2.0, allowing commercial use, modification, and distribution subject to its terms. Goose runs on the user’s machine through desktop applications for macOS, Linux, and Windows, plus a terminal client and API. Users can edit files, execute commands, and run tests. The project is active, costs $0, and does not include model usage. Block reported that about 12 specialists built the original system over nine months.

A Goose session starts in the desktop app or CLI after the user selects a language model provider. Goose sends the request and tool descriptions to that model, executes returned tool calls, and sends results back for another turn. Built in and external MCP extensions can add shell commands, file operations, web access, memory, browser automation, and application specific actions. Users can save sessions, resume them across desktop and CLI, create reusable recipes, and connect remote MCP services. Permission modes include automatic execution, approval before actions, conversation without tools, and smart approval. Permissions can also be set for each extension with allow, deny, or approval behavior. Goose supports hosted providers and local Ollama models.

Goose separates its interface, agent loop, and connected extensions. MCP provides the extension interface, while the selected provider supplies model responses and tool call requests. Documentation lists Anthropic, OpenAI, Google, Ollama, OpenRouter, Azure, and Amazon Bedrock among supported providers. Ollama can provide local inference, although speed and capability depend on the chosen model and hardware. Goose can also act as an Agent Client Protocol server and delegate to external ACP agents such as Claude Code or Codex. The project added provider configuration updates, session tools, extension controls, recipes, automatic context compaction, and security prompt settings during 2025 and 2026. Goose also joined the Linux Foundation’s Agentic AI Foundation. Block launched Goose on January 28, 2025 as an open framework for connecting language models to real world actions.

Goose software costs $0 and has no paid Goose subscription. Users provide an API key, an existing supported provider subscription, or a local model. Hosted providers may charge for tokens, requests, or subscription usage, and their quotas and rate limits remain separate from Goose. Ollama can avoid API token charges, but local inference requires suitable hardware and electricity. Goose lists a default maximum of 1,000 turns without user input, an automatic compaction threshold of 80 percent, and a compaction trigger of 225,000 tokens, subject to configuration and provider limits. Keys use the system keyring when available, with environment variables or file storage as alternatives. Setup requires review because extensions can run commands, edit files, and access external services.

Goose fits developers automating code changes, tests, shell workflows, and repeatable MCP recipes, along with teams requiring Apache licensed software, local execution, provider choice, and configurable permissions. It is less suitable for people seeking a zero setup hosted coding assistant, teams unable to review commands and extension permissions, or users wanting one bill that includes model usage. Continue and Gemini CLI are closer terminal alternatives, while Claude Code and Cursor offer more managed coding experiences. Compared with Gemini CLI, Goose puts more emphasis on model independence, local deployment, desktop and CLI access, MCP extensions, and configurable autonomy. Choose Goose when those priorities matter more than bundled inference and simpler onboarding. Developers and technical teams wanting local, open source agentic coding should pick Goose.