Goose is an open-source AI agent, created at Block, that runs on your own machine as a desktop app, CLI and API: it does not just suggest code, it reads and writes files, runs terminal commands, launches tests and installs dependencies following your natural-language instructions. Unlike an autocomplete, Goose reasons about the whole task and chains the actions needed to solve it, using the model provider you choose. In this guide you will see what Goose is, how to install the CLI and the desktop app, how to choose the model, how to extend it with MCP extensions and how it differs from other agents that work locally. The same explanation is available in Spanish.

Key takeaways

  • Goose is a general-purpose AI agent, engineering-focused, that runs on your machine (macOS, Linux and Windows) as a desktop app, CLI and API. It is open source under the Apache 2.0 licence and has over 51,000 stars on GitHub.
  • It was born at Block, the parent company of Square and Cash App, and launched on 28 January 2025; in 2026 its governance moved to the Linux Foundation’s Agentic AI Foundation. Version 1.43.0 is dated 14 July 2026.
  • It works with more than 40 providers (Anthropic, OpenAI, Google, Ollama, OpenRouter) under a "bring your own key" model: you pay only for model usage, not for the tool.
  • It is extended with more than 70 extensions through the Model Context Protocol (MCP), with official servers for GitHub, PostgreSQL, Slack and Jira.
  • It is an agent that acts: it reads and writes files, runs commands, launches tests and installs dependencies. According to Block, around 60% of its nearly 12,000 employees use it every week.

What is Goose?

Goose is an open-source AI agent that runs entirely on your own machine. That is the key difference from a plain autocomplete: it does not stop at proposing the next line, it takes a task in natural language, breaks it down and carries out the actions needed, reading and writing files, running terminal commands, launching tests and installing dependencies. It is, in essence, an AI agent that lives in your terminal and on your desktop.

It was created by Block, the parent company of Square and Cash App, as the first project of its open-source office. It launched on 28 January 2025 under the codename "goose" and was built in collaboration with Anthropic around the Model Context Protocol. In 2026 its governance moved to the Agentic AI Foundation, within the Linux Foundation, so it no longer depends on a single company. Its GitHub repository[1] (transferred to the Agentic AI Foundation’s organisation after the governance handover) has over 51,000 stars, more than 360 contributors, and is written mostly in Rust, which makes it light and quick to start.

Dhanji Prasanna, Block’s chief technology officer, summed up the goal when the project was released: the aim is to help engineers "reduce toil and give people time back to be their most creative selves". It is not an empty promise: Block itself says that around 60% of its nearly 12,000 employees use Goose every week, and not only in engineering, also in sales, design and product.

Installing the CLI and the desktop app

Goose comes in two flavours that share the same engine: a CLI for the terminal and a desktop app with a graphical interface. To install the CLI on macOS or Linux, download and run the official script with curl:

curl -fsSL https://github.com/aaif-goose/goose/releases/download/stable/download_cli.sh | bash

On macOS you can also use Homebrew with brew install block-goose-cli. The desktop app is downloaded from the project website, with packages for macOS (Silicon and Intel), Linux (DEB, RPM and Flatpak) and Windows. The app and the CLI share configuration, so you can switch between them depending on whether you prefer the mouse or the keyboard.

The first time you start the CLI it is worth launching the configuration assistant, which guides you to choose a model provider and paste the key:

goose configure            # choose model provider and credentials
goose session              # start an interactive session
goose session -r           # resume the last session
goose session -n refactor-api   # open a named session

Choosing a model provider

Goose includes no model: you decide which brain to give it, just like in other open agents such as OpenAI’s Codex CLI. It supports more than 40 providers, from commercial ones (Anthropic, OpenAI, Google Gemini, OpenRouter, Azure, AWS Bedrock) to models you run locally with Ollama so you can work without sending your code to an external service.

The choice of model matters more than in an autocomplete, because Goose relies heavily on tool calling to act. Its documentation acknowledges that it works best with the Claude 4 models, which are very strong at tool use, although any model with good function support will do. The rule of thumb is the same as with any agent: the better the model reasons, the less Goose derails on long tasks. For serious work a frontier model is best, keeping local models for bounded tasks or for privacy.

Because it works with the "bring your own key" model, you do not pay for Goose: you only pay for the usage of the provider you choose. With a cheap model or with models running on your machine, the spend can stay very low per month; with a frontier model for intensive work, it rises with usage.

Extensions and MCP

The real power of Goose lies in its extensions. An extension gives the agent a new capability (reading a database, opening an issue, searching the web) and they all connect through the Model Context Protocol (MCP), the open standard Block helped drive alongside Anthropic. There are more than 70 extensions available today, and Block maintains official servers for common tasks such as managing GitHub issues, querying PostgreSQL, sending Slack messages or working with Jira.

By building on MCP, Goose benefits from the whole ecosystem of protocol servers: any MCP server that works with Claude or with Cline (a VS Code extension that adds an autonomous coding agent to the editor) also works with Goose. And if you need a tool that does not exist, you can build your own server; we cover it step by step in building your own MCP server. Adding an extension is as simple as enabling it in goose configure or in the desktop app settings.

Goose versus other local agents

Goose competes with a wave of terminal agents that also run on your machine. OpenAI’s Codex CLI and Google’s Gemini CLI follow a similar philosophy (an agent in the terminal, with action approval), but each gravitates toward the models of its own house. Goose’s card is neutrality: being open source and compatible with more than 40 providers and with MCP, it does not tie you to any model maker or any specific editor.

Against VS Code extensions like Cline, Goose bets on a dual experience: the CLI for those who live in the terminal and a standalone desktop app for those who prefer their own window, rather than living inside the editor. And unlike closed paid agents, it does not charge for the tool and you can audit all of its code under the Apache 2.0 licence. The trade-off is the usual one for open agents: you manage the keys and choose the model yourself, which demands a little more initial setup in exchange for transparency and control.

Frequently asked questions

Is Goose free?

Yes. Goose is free software under the Apache 2.0 licence, so you pay nothing for the tool and can read, audit and modify all of its code. What does cost money is the language-model usage, because Goose works with your own API key and each task consumes tokens. With a cheap model or with models you run on your own machine via Ollama, the spend can stay very low per month; with a frontier model for intensive work, it rises with usage.

Can I use Goose offline with a local model?

Yes. Goose supports Ollama as a provider, so you can connect an open model running on your own machine and work without sending your code to an external service. Bear in mind that Goose relies heavily on tool calling, so it is worth picking an open model with good function support; local models tend to reason worse on long tasks than a frontier model.

What sets Goose apart from Cline or the Codex CLI?

Cline lives inside VS Code as an extension; the Codex CLI and the Gemini CLI are terminal agents tied to a model maker. Goose is more neutral: it runs as a CLI and as a standalone desktop app, is open source and works with more than 40 providers and with MCP extensions, without tying you to any model or editor. In exchange, you manage the keys and choose the model yourself.

Conclusion

Goose is one of the strongest open-source bets for bringing a coding agent to your own machine: it was born at Block, is now governed by the Agentic AI Foundation, runs as a CLI and as a desktop app, and extends without limits thanks to MCP. Its greatest virtue is neutrality, because it works with the model and the tools you choose without tying you to anyone. The natural next step is to install it, connect a provider and try it on a real task, and to compare it with OpenAI’s Codex CLI or Google’s Gemini CLI to decide which fits your way of working best.

Sources: [1] Official Goose documentation[2], [2] Goose on GitHub[1], [3] Block introduces codename goose[3], [4] Block launches the Goose agent, Fortune[4].

Visitas: 0

Sources

  1. GitHub repository
  2. Official Goose documentation
  3. Block introduces codename goose
  4. Block launches the Goose agent, Fortune

Route: AI Coding Agents