Goose review

Goose is a general-purpose AI agent that runs on your machine, rather than a hosted-only agent service. Its user-facing surfaces are a native desktop app for macOS, Linux and Windows, a CLI for terminal workflows, and an API for embedding it elsewhere. The project says it is built in Rust; that describes Goose’s implementation, not a requirement to write integrations in Rust.

In a stack, Goose sits between a user or application and an LLM provider, with MCP extensions connecting it to tools and data sources. It supports API keys and eligible Claude, ChatGPT or Gemini subscriptions through ACP, according to the homepage. The Goose Development Kit provides runtime components for building agents across native apps, servers and browsers, with saved-state resumption. Recipes capture workflows in YAML, including instructions, extensions, parameters and subrecipes. This makes Goose an agent runtime and integration layer, not a model or a managed inference service. Its stated scope includes research, writing, automation, data analysis and coding.

The homepage describes an API for embedding Goose, but does not name an API style, endpoint specification or SDK language. It names Rust as the language Goose is built in, and the Goose Development Kit as the runtime components for building agents in native apps, servers and browsers. Confirm SDK support and API details in the vendor’s docs before choosing an integration path.

For terminal workflows, Goose provides a CLI with an install command on its homepage. The desktop app runs on macOS, Linux and Windows. Recipes use portable YAML, and the homepage says they can be shared, run in CI and include extensions, parameters and subrecipes. MCP is the tool-and-data integration standard; Goose lists more than 70 documented extensions. Goose also works as an ACP server for clients including Zed, JetBrains and VS Code, and can use ACP agents such as Claude Code and Codex as providers. MCP Apps can display interactive buttons, forms and visualizations inside Goose Desktop. Check the docs for current setup instructions and compatibility.

The Open Source plan costs $0 per month to run and includes the Apache 2.0 software. The pricing information lists bring-your-own LLM provider or API keys; the homepage also describes eligible Claude, ChatGPT and Gemini subscriptions through ACP. There is no listed Goose free-trial period, and annual billing is not published.

The $0 figure is the software price, not a bundled allowance for model inference. With API keys, usage costs depend on the provider and model you select; Goose’s pricing information does not publish a combined per-request rate. With local models, inference runs on your own machine, so hardware and power are your responsibility. The listed plan has no published seat or usage limits, and the vendor does not publish Goose rate limits, overage charges or add-ons in the published pricing details. That means there is no documented Goose usage bill to extrapolate, but provider charges can still grow with model use. Before a production rollout, verify current provider pricing and any limits that apply to your chosen model or account.

Goose is open source under the Apache 2.0 license, and its homepage points developers to GitHub to browse, fork and contribute to the project. It runs on your machine, with desktop and CLI options, rather than requiring a Goose-hosted runtime. The project also describes a Rust implementation and a Development Kit for embedding its runtime components.

Provider choice and open standards can reduce dependence on one model vendor: Goose supports multiple providers, MCP extensions and ACP connections. Recipes are portable YAML files that can be shared and run in CI. Those are useful portability points, but they are not a complete data-export guarantee. The homepage does not describe a session-data export process or migration procedure. Before adopting Goose for durable workflows, ask how conversations, configuration and other state can be backed up and moved, and check whether your extensions rely on provider-specific behavior.

Goose lists prompt-injection detection, tool permissions, sandbox mode and an adversary reviewer as security controls. Those controls do not replace a review of tool access, secrets handling and the blast radius of actions in your environment. The homepage does not publish rate limits, availability commitments, scaling guidance or compliance certifications. Confirm those requirements, plus provider-side limits, before making Goose a production dependency.

Reddit reports are mixed on model behavior and the desktop experience. In one LocalLLaMA thread, users describe variability in task completion across open models; another commenter reports a macOS keychain-access prompt that raised security concerns. A separate GUI user asks for projects available in the CLI to be added to the GUI, while another user recommends Goose for general MCP tasks. These are individual reports, not controlled evaluations, but they point to practical checks: test your chosen model and tools, review credential access, and verify that the UI covers your workflow. Do not assume one model’s results transfer to another.