Goose vs Claude Code (2026): goose is not Block's agent any more

Quick answer

Goose and Claude Code are both terminal-first coding agents, and in 2026 the comparison almost everyone is running is built on a premise that expired. Goose is not Block's project any more. Block donated it to the Agentic AI Foundation at the Linux Foundation on April 7, 2026 (goose docs, 2026), and Anthropic donated the Model Context Protocol to that same foundation. Goose's extensions are MCP servers. The two are now sibling projects under one neutral steward, so the real choice is about harness behaviour, not about which vendor you are siding with.

I went in wanting a feature table. I came out with a governance correction, because the thing that changed most about goose in 2026 was not a feature.

The repo moved and the comparisons did not

I went to clone goose from the path a ranking article gave me and got redirected. The project now lives at aaif-goose/goose, not block/goose (repository, 2026). 53.2k stars, still Apache-2.0. The documentation moved too, off block.github.io/goose and onto goose-docs.ai.

The GOVERNANCE.md in that repo is blunt about it: goose was "Founded by Block and now stewarded by AAIF", operating as a Series of LF Projects, LLC. Day-to-day authority sits with three to seven Core Maintainers, most decisions happen by consensus in pull requests and discussions, and architectural changes need a majority of Core Maintainers after a week of community review.

So I checked what the pages ranking for this query actually say. I fetched them and counted terms rather than skimming:

Four of the five I could verify carry no mention of the change at all. One is still sending readers to a repository path that has moved. The single article that does mention it is not a versus piece.

I am not scoring anyone for this. Governance footnotes are exactly the kind of thing a comparison post skips. But one of these pages is titled around goose being "Block's OSS Agent" and that is now the wrong frame for the argument it is making.

Anthropic is in the same foundation

This is the part that reorganised my thinking.

The Agentic AI Foundation was announced on December 9, 2025, and it did not launch with one donation. It launched with three: Anthropic's Model Context Protocol, Block's goose, and OpenAI's AGENTS.md (foundation announcement, 2025). Anthropic, Block, OpenAI, AWS, Bloomberg, Cloudflare, Google and Microsoft all sit at the Platinum tier.

Read that against the framing the SERP is saturated with. The open-source challenger and the closed incumbent co-founded the same neutral standards body, at the same time, and each handed it a core piece of their agent stack.

That does not make the products identical. It does mean "which vendor am I backing" is a weaker reason to choose between them in 2026 than it was in 2025, because on the layer where lock-in would actually bite, they deliberately pooled.

The extensions are MCP servers, which is the real punchline

I had assumed goose's extension system was its own thing that happened to look like MCP. It is not. The docs say extensions "are based on the Model Context Protocol (MCP)", and then go further than I expected:

> goose's built-in extensions are MCP servers in their own right. If you'd like to use the MCP servers included with goose with any other agent, you are free to do so.

Claude Code speaks MCP as a first-class citizen (Claude Code MCP docs, 2026). So the integration layer people describe as the dividing line is the shared part. An MCP server I wrote for one runs under the other. I have run an MCP server in production for a month and the thing I would have had to rewrite to switch agents turned out to be the thing I did not have to touch.

Goose ships several extensions on by default, including Developer, Computer Controller, Memory, and a Tutorial one. You toggle them through goose configure or grab one directly:

goose configure
goose mcp developer

So I pointed goose at Claude

The other assumption worth killing is that goose means giving up Claude models. Anthropic is a supported provider in goose configure, so you can run goose on the same model Claude Code is running.

That is the only honest way I know to compare two harnesses. Hold the model still and the difference you observe is the harness, not the weights.

goose configure
goose session
goose session -r

Install is a single script, and the release path is the one that gives away the move:

curl -fsSL https://github.com/aaif-goose/goose/releases/download/stable/download_cli.sh | bash

I ran one task through both on the same Anthropic model: add a small export endpoint to an existing service, with tests. One task, one afternoon, so treat this as an anecdote and not a benchmark. Two things were consistently different, and both follow from design rather than luck.

Goose reached for tools earlier and more literally. With the Computer Controller and Developer extensions live, it treats "go look at the thing" as a normal step. Claude Code did more reading of the repository before touching anything. Neither is strictly better. On a codebase I knew well, the reading felt like latency. On the part where I was wrong about how an existing helper worked, the reading is what caught it.

The second difference is scope of ambition. Goose is not a coding agent that grew extensions, it is a general agent that codes, and it behaves like one. That is a genuine advantage when the task leaves the repository, and a genuine liability when you wanted it to stay in the repository.

What the foundation move does not fix

I want to be careful here, because it would be easy to let a governance finding do work it cannot do.

Donating a project to a foundation does not make it better at writing code. It does not change model quality, harness polish, or how good the diffs are. It is not a promise of maintenance velocity either. Plenty of foundation-hosted projects are quiet. The GOVERNANCE.md still names a Block engineer as the tie-breaker when Core Maintainers deadlock, which is normal for a recently donated project and is also not what "vendor-neutral" sounds like when you say it out loud.

And Claude Code remains the more finished product for the narrow job of editing code in a repository. That was true before April 2026 and the donation did not touch it.

What the move actually changes is the shape of the risk you are taking. In 2025, choosing goose meant betting that one payments company would keep funding a side project. In 2026 it means betting on a Linux Foundation project with published governance and a named maintainer group. That is a different, smaller bet, and it is the part no comparison told me.

Which one I keep

Both, and not as a diplomatic answer.

Claude Code stays my default for work inside a repository. It is better at the specific loop of read, propose, edit, run tests, and I do that loop most days.

Goose earns its place for work that starts in a repository and ends somewhere else, because the extension model is honest about the agent leaving. When I need the same capability under both, I write it as an MCP server once, which is the practical dividend of the thing this whole article is about.

The framing I would retire is "open source versus Anthropic". As of 2026 the open protocol underneath both is an Anthropic donation, sitting in the same foundation as goose. If you are choosing between them, choose on how the harness behaves on your work. The vendor-loyalty version of this question got settled by the vendors.

Checked against first-party sources on August 22, 2026. If you are reading a goose comparison written before April 2026, check the repository path before you trust the rest of it. And if a comparison tells you a harness is locked to one vendor's models, check that too. It is increasingly not the case.