Goose Was Supposed to Write Your Code

Dorsey's company ships an open-source AI agent called Goose — runs locally, uses tools, executes code, browses the web, the whole thing — and the pitch was basically: it's for developers, it lives in your terminal, it helps you build software.

That part's going fine. But something else is happening.

People are using it to manage their email. Organize research. Summarize documents. Plan trips. The coding agent is quietly becoming a general-purpose agent because, it turns out, "run arbitrary tools against a model that can reason" doesn't care what your job title is.

This is the thing nobody wants to say out loud: most AI agents are just the same infrastructure wearing a different costume. You call it a "coding agent" and developers use it. You call it a "research agent" and analysts use it. You call it an "executive assistant" and you can charge more. Goose shipped without the costume and people are discovering that the underlying skeleton does whatever you point it at.

The GitHub repo is picking up stars fast right now — the kind of trajectory that means a project tipped from "people who follow this space" into "people who found it on a list somewhere." That second group doesn't know what the intended use case was. They just see a local agent that can use tools and they start plugging in their actual problems.

The irony is Block framed this as a developer tool, which gave it credibility and an early adopter base that actually knows how to configure things — and that base then built the on-ramps for everyone else. Same way git got popular because programmers used it, then writers found out about it, then everyone was using it for everything.

Goose is probably not unique in having this property. It might just be the one that's visible right now, this week, in this particular moment when people are bored of the hosted options and want something that doesn't phone home.

Local matters more than anyone at the big labs wants to admit. Running something on your own machine — even if it's slightly worse, even if the setup takes an hour — removes a whole category of anxiety that turns out to have been suppressing use cases. People will do more with a tool they trust to stay on the box. That's not rational, but rationality was never the constraint.