Goose

Goose is Block's answer to the question: what if your AI agent lived on your machine instead of someone else's cloud? It's a command-line agent that plans multi-step tasks, executes them, and iterates when things go wrong. The autonomy shows up in how it chains actions: ask it to refactor a codebase, and it will scan files, propose changes, run tests, fix breakages, and commit the result without you babysitting each step. I've used it for cleaning up legacy Python projects. Point it at a directory, describe the refactor, and it will work through files methodically. It caught edge cases I would have missed and fixed its own mistakes when tests failed. The local execution matters: it reads your filesystem, runs shell commands, and respects your environment without uploading your code to a third party. For anyone working with proprietary codebases, that's the entire point. The MCP integration is the clever bit. Goose doesn't ship with every possible capability baked in. Instead, it connects to MCP servers for specific tasks: filesystem access, database queries, API calls. You configure which servers it can use, and it figures out when to invoke them. This makes it extensible without bloating the core agent. If you need it to interact with a new service, you add an MCP server rather than waiting for Block to ship an update. Where it stumbles: the command-line interface is spartan. No GUI, no progress visualisation beyond text logs. If a task goes off the rails, you're reading through terminal output to diagnose it. The planning is decent but not magical. On ambiguous tasks, it will pick a direction and commit rather than asking clarifying questions. I've had it refactor code in ways that were technically correct but missed the intent. Compared to Aider, Goose is broader but less polished for pure coding. Aider understands Git workflows and code context better. Goose tries to be a general assistant, which means it's competent at coding, data analysis, and automation but not the sharpest tool for any single job. The trade-off is flexibility: Goose will draft documents, query databases, and automate workflows in the same session. The open-source model is a double-edged sword. You can audit the code and run it offline, but you're also responsible for setup and troubleshooting. No support team, no polished onboarding. If you're comfortable with command-line tools and want an agent that stays on your machine, Goose delivers. If you want something that just works out of the box, look elsewhere.