Anyone who has used a conversational programming assistant knows the ritual. You ask for a function, the answer is articulate, and then you copy it into the editor yourself, install the dependency, run the tests, and come back with the error message so the assistant can try again. Each round trip takes back some of the time the assistant just saved, and after a few of them the net gain is hard to see.

Goose removes that step. It does not answer questions and stop; it works on your machine. It reads files, changes code, runs commands, looks at the output and continues from there until the task is done. It was built inside a financial technology company and released as open source, then handed in the spring of 2026 to a foundation that looks after open infrastructure, under a permissive licence, with tens of thousands of watchers and hundreds of contributors.

Not only a coding assistant

The project calls itself general purpose, and means it in a plain way. Writing code is one of the things it does, alongside looking things up, drafting documents, working through a data set and running an automated sequence. Ask it a question in a terminal and it plans the steps, calls the tools it needs, checks what came back, and adjusts. It behaves less like a tool you operate and more like an intern who is always available and never needs the context explained twice.

There are three ways in. A native desktop application for the three common platforms, for people who would rather click than type. A command line tool, for people who would rather not. And an interface for embedding the agent in somebody else's product. The core is written in Rust, which shows in how quickly it starts and how little it asks of the machine.

Seventy-odd extensions, and the freedom to add more

Capability comes from extensions. The project builds on an open standard for connecting a model to external tools, and the community has contributed well over seventy ready-made ones: databases, browsers, code hosting, cloud storage, assorted services. Asking it to check a document while it edits, call an internal endpoint, or drive a browser to verify that a page actually renders no longer means switching applications.

It can also expose itself as a server that the popular editors understand, so the same agent appears inside the interfaces people already use. That detail matters more than it sounds: an agent that lives in a separate window is an agent people forget to use.

Any model, including one you already pay for

No vendor is favoured. More than fifteen providers can be attached, from the large commercial ones to a model running locally on your own hardware, where nothing leaves the machine at all. And for people who already hold a subscription somewhere, an open protocol lets that account be used directly, without obtaining a second key.

Everything runs locally and data does not leave the computer by default, which is the difference between a tool that can be used on confidential work and one that cannot. The project is also unusually explicit about safety: detection of prompt injection, per-tool permissions, a sandbox mode, and a reviewer that watches for dangerous actions before they happen.

Saving a routine as a recipe

Some procedures repeat every week: run the tests, collect the failures, fix them one at a time, run the tests again to confirm, then commit. Goose lets that be written down once as a recipe in a plain text file. Afterwards a single sentence triggers it, and the agent follows the sequence on its own, shares it with a colleague, or runs it inside a build pipeline. A collection of recipes has already accumulated in the community, close enough to the usual cases that adjusting one is faster than writing from scratch.

Larger tasks can be split across parallel sub-agents: one reviewing code while another researches a question, without the two polluting each other's context and without the main conversation filling up with logs. It is the difference between hiring one assistant and running a small office.

Who it is for

Developers and teams who want coding, research and automation in one place, and who cannot send the repository somewhere else, will find this the most practical arrangement available: free, open, model agnostic, extensible. What it removes is not thinking but the copying and pasting that used to sit between thinking and result.