Goose started at Block and has since moved to the Agentic AI Foundation at the Linux Foundation, which is the detail I would lead with if I were writing their announcement. Open source projects that live inside one company are always one strategy change away from abandonment, and agent tooling is infrastructure that people build workflows on top of, so governance stability is not a nicety. Handing stewardship to a vendor-neutral foundation with 500-plus contributors behind it is the difference between a tool you can depend on and a tool you are borrowing. Written in Rust, which shows up as a desktop application that starts fast and does not feel like a browser wearing a costume, plus a CLI and an API.
It runs on your machine rather than in someone's cloud, which is the correct default for an agent that will read your files and run your commands, and it works across 15-plus model providers, so you are not locked to whoever is currently winning. The extension catalogue, 70-plus covering databases, APIs, browsers, GitHub and Google Drive, is the practical breadth that makes it useful for things beyond coding. Research, data analysis, writing, automation. It is a general agent and it means it.
Recipes are the feature I have come to think is the most important and the least discussed. A recipe is a portable YAML file describing a workflow. That sounds mundane and it changes the nature of the thing. Without it, using an agent is a series of prompts that live in your head and your history, and the knowledge of how to make it do a useful job is not transferable.
With recipes, a workflow becomes an artefact you can version, review in a pull request, share with a colleague and improve over time. This moves agent usage from craft to engineering, and it is the single practice I would push any team adopting agents to adopt alongside them. The security posture deserves credit because it is unusual. Prompt injection detection and a sandbox mode, shipped as features rather than mentioned in a FAQ.
Most agent projects hand you something that will execute arbitrary commands based on text it read from the internet and consider the risk your problem. Goose at least ships mitigations. Read the security documentation properly before you point this at anything that matters, because a local agent has local blast radius and the failure modes are not theoretical. The honest weakness of a general purpose agent is that it is second best at most specific things.
For coding specifically, dedicated tools are better, because they have been shaped around that workflow for years. Goose's argument is breadth, that you have one agent that can move between a database, a browser and a document rather than three tools that each do one. Whether that is worth the trade depends entirely on how varied your work is. If you code all day, use a coding agent.
If your day is genuinely mixed, this is the better shape. Costs are worth saying plainly since the tool is free. You pay for model API calls, and an agentic loop that explores, retries and self-corrects makes many more calls than a chat session. People who assume free means free get a surprising bill in week one.
Pick a cheaper model for routine steps, set limits, and watch usage until you have a feel for it. Four point three. Free, vendor-neutral, thoughtfully built, and honest about what it is.